The AI Contract Workflow: Draft, Send, Track and Close Agreements With ChatGPT or Claude

TL;DR: You can run an entire agreement lifecycle through an AI assistant: draft, review, send for signature, watch what the client actually does with the document, follow up, and collect. The prompts that work are full business briefs, not vague requests. Human review of the operative clauses (parties, amounts, dates, termination, liability) is not optional, and AI does not replace a lawyer. What closes the loop is [okdoc](/): documents are created as drafts you approve, and Track shows opens, dwell time, time on each page, scroll depth, a click heatmap, and a live WhatsApp alert the moment your client opens it. Track is free on every plan, including FREE ₪0. API, webhooks and the MCP connector are an AGENCY ₪499 entitlement.

Most people who try to "do contracts with AI" stop at step one. They ask ChatGPT to draft an agreement, get some text, and then go back to working by hand: copy into a word processor, format, export a PDF, log into an e-signature tool, upload, place fields, send, wait.

That is not a workflow. That is a draft with manual labor wrapped around it.

This guide is about the whole thing: how an agreement goes from an idea to money in your account, with the AI doing the part it is genuinely good at and you doing the part you must never hand over. If you have not connected okdoc to your assistant yet, that setup is covered separately in signing documents directly from ChatGPT. Here we assume the plumbing exists, or that you are working from the okdoc interface, and we talk about the method itself.

The five beats of a deal, and where AI fits

Every client agreement moves through the same five beats, whether you think about it that way or not:

Create → Send → Track → Sign → Close.

The common mistake is treating this as a signing process. Signing is one beat of five, and it is usually not where deals stall. Deals stall between Send and Sign, in the gray zone where you have no idea whether the client opened it, read it, hesitated, or forgot.

Here is an honest split of what AI contributes and what it does not:

| Beat | What AI does well | What it does not do | |---|---|---| | Create | Structure, language, standard clauses, variants | It does not set your commercial or legal policy | | Send | Assembling recipients, signer order, payment terms | It does not decide when you should send | | Track | Summarizing statuses, spotting who is stuck | It cannot see data you never collected | | Sign | Nothing. Deliberately | It never signs on anyone's behalf | | Close | Drafting follow-ups, reminders, revenue summaries | It does not collect the money for you |

That "Sign" row matters most. An AI assistant does not sign. It prepares, sends and reminds. The signature itself is always made by the signer, on their own device, after verification. That is a deliberate design principle, not a limitation someone will lift in the next release.

Beat 1: Drafting, and what a prompt that actually works looks like

The difference between a usable first draft and generic legal-sounding filler is how much context you loaded into the first request. A weak prompt sounds like "write me a service agreement." A strong one looks nothing like that.

The brief formula: seven elements

A good contract prompt contains seven things. If one is missing, the model will invent it, and you will find out later:

• 1. Who the parties are. Full legal names, entity type, state of formation, addresses. Not "the client" and "the company."
• 2. What exactly the deliverable is. Not "marketing" but "management of paid social campaigns, up to four new creative variants per month, plus one monthly performance report."
• 3. How much and when. Amount, currency, whether tax is included, payment schedule, what happens on late payment.
• 4. Term and start date. Effective date, length, whether it auto-renews, and how someone gets out of it.
• 5. What is explicitly out of scope. This clause prevents more disputes than any other single clause.
• 6. Termination. Who can terminate, on how many days' notice, and what happens to money already paid.
• 7. Tone and length. "Plain business English, two pages" and "comprehensive and detailed" produce two completely different documents.

Prompts that work

A project proposal:

"Draft a proposal from [My Company LLC, a Delaware LLC, address] to Dana Cohen at Network Inc. Deliverable: an 8-page WordPress marketing site, mobile responsive, with Google Analytics installed. Price $12,000. Payment: 40% deposit on signature, 60% on delivery. Timeline: 6 weeks from receipt of all content. Out of scope: copywriting, photography, domain and hosting purchase. One revision round included, additional rounds billed at $125/hour. Proposal valid 14 days. Plain business tone, not heavy legalese, under two pages."

A monthly retainer agreement:

"Draft a monthly retainer agreement for marketing services. Parties: [My Company LLC] and Alan Levy of Pizza Center Inc. $2,500 per month, invoiced on the 1st, net 15. Ad spend is paid directly by the client to the platforms, never through my account. Term: first 3 months are committed, then month to month with 30 days' written notice by either party. No auto-renewal, I want the client to actively renew. Include a clause that all creative assets and ad accounts remain the client's property at termination. Include a mutual confidentiality clause. Governing law: Delaware."

An NDA before a first meeting:

"Draft a mutual NDA between [My Company LLC] and Acme Ltd for exploratory discussions about a technology partnership. Confidentiality period 3 years from disclosure. Broad definition of confidential information, with the standard carve-outs: already public, lawfully received from a third party, and independently developed. Governing law New York, venue New York County. One page."

Notice the pattern. You describe a business decision you already made, and the AI expresses it. Not the other way around.

The trick that changes output quality: give it your last agreement

The fastest way to get a draft in your voice is not to describe your voice, it is to show it:

"Here is the last agreement I signed with a client. Keep the same structure, the same section breakdown and the same tone, and only replace the commercial terms per this brief: ..."

This beats any abstract instruction about "professional tone." It also has a second benefit: clauses that already survived your review stay exactly as they are, which shrinks the surface area of things the model invented from scratch.

If you work in okdoc, the obvious next move is to save that document as a template and keep sending from it. One good template saves more time than any brilliant prompt.

What you should never let AI invent

Some things must not be left to a language model's guess, no matter how confident it sounds:

• Amounts, percentages and payment dates. Never leave these open. If you did not say it, the model fills in something plausible, and plausible is not correct.
• Liquidated damages and liability caps. These are numbers with real legal and financial consequences. You decide them.
• Governing law and venue. Especially in cross-border or multi-state deals.
• Regulated-industry clauses. Healthcare, financial services, insurance, real estate. There are requirements there that no generic template contains.
• Entity names, EINs, addresses and dates. A language model reproduces patterns, it does not remember facts about your business. If you did not paste the number, it may produce a correctly formatted but entirely fictional one.

The simple rule: AI is good at structure and phrasing, not at decisions. Anything that is actually a decision, you make.

Beat 2: Review, and what a human has to read personally

This is the beat people skip, and it is the only one you genuinely cannot.

An AI draft is a starting point, never final legal text. A language model produces text that looks right. "Looks right" and "is right" are two different things, and in a contract the gap between them costs money.

The operative-clause checklist

An "operative clause" is any clause somebody could sue on. These are the ones to read word by word, with your own eyes, before the document leaves your hands:

• Parties. Exact legal names, correct entity type, and who signs on whose behalf. An agreement with "Pizza Center" when the entity is "Pizza Center Management, Inc." is a real problem on the day there is a dispute.
• Consideration. The amount, the currency, whether tax is included, and exactly what earns the payment.
• Payment terms. Net how many days, counted from what (invoice date? delivery?), and what happens when it is late.
• Deliverables and scope. What is in, and above all what is out. An empty "out of scope" section is an invitation to scope creep.
• Schedule. Start date, milestones, and what happens when the client is the one causing the delay.
• Termination. Who, how, on what notice, and what happens to money already paid.
• Auto-renewal. If it exists, what the cancellation window is. This is the clause the largest number of businesses discover too late.
• Limitation of liability. The cap, the carve-outs, and what is actually covered.
• Intellectual property. Who owns what at the end, and what happens if you were not paid.
• Confidentiality. Duration, exclusions, and what happens to the material at termination.
• Governing law, venue, and dispute resolution.

Three numeric checks that take one minute

Before you approve a send, run these three. They catch most of the embarrassing errors:

• 1. The name check. Every name, entity and address in the document. Especially after duplicating an earlier document, because a leftover name from the previous deal buried in section 7 is the single most common mistake in the world.
• 2. The number check. Every amount appears at least twice in a typical document, in the price section and in the payment schedule. Confirm they agree with each other.
• 3. The date check. Effective date, end date, proposal expiry. Do the arithmetic in your head. Do not trust the model's date math.

Using AI to check the AI

Worth doing, but in a separate conversation and with a narrow request. Not "is this good?" but:

"Here is a contract. Do not rewrite it. Return a table only: every obligation of mine, every obligation of theirs, every date, every dollar amount, and every clause that lets the other side exit or charge me more. Flag any internal contradiction or any cross-reference to a section that does not exist."

This output is genuinely useful, because it forces you to look at the document as a list of commitments rather than as flowing text. But be clear about what it is: reading assistance, not approval. The AI does not know what is normal in your industry and does not know what you agreed to on the phone.

When to involve a lawyer

Without being precious about it, there are situations where an AI draft is the starting point for a conversation with counsel, not a substitute for one:

• Agreements you will reuse at volume, like customer terms or a standard MSA you will put in front of hundreds of clients. Pay once for language that has been reviewed, then reuse it forever.
• Amounts it would hurt to lose.
• Partnerships, equity, investment.
• Regulated industries.
• Anything with a counterparty in another country and foreign governing law.
• Anything you still do not understand after reading it twice.

The efficient pattern: AI produces the draft and the question list, the lawyer reviews and corrects, and you save the result as a template and use it repeatedly. One legal bill, years of value.

Beat 3: Sending, and the decisions locked in at that moment

This is where a draft becomes a real document in front of a real client. Several decisions get fixed here that are hard to change afterward.

Draft first, send second

In okdoc, when an AI assistant creates a document through the connector, the default is a draft. The document is created in your account, you get a dashboard link, and you open it, read it, fix it and send it yourself. Immediate sending is an explicit choice that has to be requested.

That is not a design oversight, it is the single most important safeguard in this entire workflow. A document that reaches a client without a human reading it is exactly the scenario that turns a nice automation into an incident.

Good prompt habit: "Draft it, but do not send. Give me the link to review."

What gets decided at send time

• Who signs, and in what order. One signer or several, and with several: sequential (one after another) or parallel (everyone at once). Sequential fits a hierarchy, for example the client signs and then your principal countersigns. Parallel is faster.
• Identity verification with a one-time code. You can require the signer to enter a code emailed to them before they can sign. Always worth it on larger amounts, or when you have never met the counterparty.
• An offer timer. You can attach a countdown to a proposal and decide what happens at expiry: lock the document or just notify. It works, but only if the deadline is real. A deadline you extend every time teaches the client to ignore it.
• Sign and pay on the same screen. You can charge a card at signing through a connected payment provider, or show a payment link or bank transfer instructions. More on this on the proposals and quotes page.
• The delivery channel. Email or WhatsApp. WhatsApp is opened faster, consistently.

What counts as a "deal"

An operational detail worth knowing: okdoc's monthly quota counts documents actually sent. Drafts never count. You can generate ten versions of the same proposal, agonize, delete nine, and only pay for the one that went out. FREE ₪0 includes 5 deals a month, CLOSE ₪199 includes 50, and AGENCY ₪499 includes 200. Full details on the pricing page.

Beat 4: Track, the part no AI tool will give you

This is the real gap, and it is worth being blunt about it: AI can write the document. It has no idea what your client did with it.

An AI assistant cannot see that your client opened the proposal at 7:40pm, read for four minutes, got stuck on the pricing table, scrolled down to the cancellation clause, and closed it. Only a system sitting on the signing page itself can see that.

That is exactly what okdoc's Track does, and it is free on every plan, including FREE ₪0. Not an add-on, not an upgrade.

What gets measured

• Opens. How many times the document was opened, and exactly when each time.
• Dwell time. How long the client actively read, not how long a tab sat open.
• Time on each page or section. Where the minutes actually went.
• Scroll depth. How far they got. Someone who stopped at 30% did not read your terms.
• Click heatmap. What they clicked on the page.
• Deal heat. One status that summarizes it: hot, viewed, abandoned, or stalled.
• A live WhatsApp alert. A message to your business number the moment a client genuinely starts reading.

Reading the signals, and what to do about each

Data is only worth something if it produces an action. Here is the translation:

Hot. Opened right now, or reopened, or read deeply. This is the moment your client is thinking about your deal. If you get to pick one time each day to pick up the phone, this is the time and this is the client.

Abandoned. Opened, barely scrolled, never came back. Something near the top of the document stopped them. Usually the price on page one, or an opening section that is too long. The right follow-up here is not "did you see it?" but a response to the actual friction: "I wanted to make sure the payment structure is clear, we can also break it into installments."

Stalled. Sent days ago, never opened at all. That is not a content problem, it is a deliverability or attention problem. Check whether the email landed in spam, then resend over WhatsApp.

A dwell spike on one page. If all the time landed on a single page, that is where the hesitation lives. A pricing page that eats four minutes means there is a question about price that nobody asked out loud.

Stopped right before the signature. Reached the bottom, did not sign. Usually that is authority (they need to ask someone) or one open question. A short message offering a five-minute call solves it far better than an automated reminder.

The combination: AI drafts it, Track tells you whether it landed

This is the heart of the method. AI gives you speed in creation. Track gives you sight. Without Track, follow-up is guesswork. With Track, follow-up is a response to something that actually happened.

And you can close the loop back into the AI. Once you know what the client did, the follow-up prompt becomes concrete:

"Dana opened the proposal twice, spent four minutes total, most of it on the payment terms section, and never reached the end. Draft a short, non-pushy follow-up message offering an alternative payment schedule and asking if tomorrow works for a quick call. Four lines maximum, no exclamation marks."

That prompt is good not because of how it is phrased, but because it contains facts.

Beat 5: Follow-up without becoming a nuisance

There are three levels of follow-up, and you should use all three in different places:

Manual reminders. You decide, you send. A connected assistant can find unsigned documents and send a reminder to the pending signer. Prompt: "What went out in the last two weeks and still isn't signed? Show me a list, and don't send anything until I approve them one by one."

Automations. Time-based reminders, alerts, and post-signature actions. That is a CLOSE ₪199 entitlement and up. Their strength is consistency. Their weakness is that they do not know your client is on vacation.

Signal-based human follow-up. This is what produces the most closed deals, and it is what the WhatsApp alert exists for.

A rule of thumb on cadence: one reminder after 48 hours, a second after a week, then switch channels or pick up the phone. Three identical automated reminders is not persistence, it is why people hit the spam button.

Beat 6: Close. Signature, sealed copy, money

Signature and audit trail. Once the document is signed, a sealed copy is produced that makes post-signature tampering detectable. There is a full event log: when it was sent, when it was opened, when it was signed. If in two years you ask yourself "what exactly did they sign, and when," the answer is there. Background on the mechanics: how electronic signatures work and what an e-signature actually means.

okdoc is designed to support legally binding electronic signatures under ESIGN and UETA. For a deeper look at the legal framework, see e-signature legality and the primer on what an electronic signature is.

Collection. If you attached sign-and-pay, the client signs and pays on the same screen. This is the single biggest lever on collection rate. The gap between "signed" and "paid" is where most of the money gets delayed.

What happens after. You can download the sealed copy, save the document as a template for next time, file it in a folder, and pull a revenue report. On AGENCY you can also connect a webhook that pushes events into your own systems: `document.sent`, `document.viewed`, `document.signed`, `payment.received`. That is what lets a signed contract automatically update a CRM or a bookkeeping system.

The other half: reviewing a contract you received

Everything above is about documents you send. The other half, and often the more consequential half, is documents somebody sends you. A vendor agreement. A platform's terms. A large client who arrives with their own paper.

This is where AI is genuinely excellent, because the task is extracting information rather than producing it.

What to ask it to look for

• Every obligation of mine, as a list. Not a summary, a list.
• Every date. Effective, expiry, notice deadlines, cancellation windows.
• Auto-renewal. Does it exist? What is the exit window? What happens if I miss it?
• Payment terms. When, for what, late fees, offsets, and whether they can withhold.
• Limitation of liability and indemnification. Who covers whom, up to what, and what is carved out.
• Termination. Who can terminate and when, and what happens to prepaid amounts.
• Exclusivity and non-compete. Anything that restricts who else you can work with.
• Intellectual property. What transfers, what stays yours, what happens to work product.
• Unilateral change rights. Can the other side change terms or pricing on notice alone?
• Venue and arbitration. Including where, and who bears the cost.
• References to external documents. "Per the policy posted at X" pulls text you have never read into your contract.
• Internal contradictions and broken cross-references.

A ready-to-use prompt for an incoming contract

"I am receiving this contract, not writing it. Go through it and return: (1) a table of every obligation of mine against every obligation of theirs, with section numbers; (2) every date and time window in the document; (3) whether there is auto-renewal and exactly what I must do to avoid getting locked in; (4) the five riskiest clauses for me, ranked by severity, with a one-line reason each; (5) for each of those, alternative language I could propose; (6) every reference to an external document that was not attached. Do not rewrite the contract and do not give legal advice, just show me what it says."

That last instruction matters. You want extraction, not opinion.

okdoc has a dedicated tool for this that returns a summary, key terms, and risky clauses with a severity rating and recommendations, and it is available through the connector as well. It costs one AI credit per analysis.

Where AI fails at reading contracts

Be aware of three systematic weaknesses:

• 1. It sees what is there, not what is missing. The riskiest clause in a contract is usually the one that was never written. No liability cap. No termination clause. No definition of what counts as "acceptance." A language model will not reliably surface an absence unless you explicitly ask it to hunt for absences.
• 2. Incorporated documents. If the contract points to an exhibit or a policy on a website, that text is not in front of the model. It will only analyze what you pasted.
• 3. Business context. It does not know that this term is unusual in your industry, or that you agreed to something different on the call. It checks text, not understandings.

Confidentiality: what actually leaves your system

This is the part most content on this topic skips, and it is precisely the part that matters.

The baseline rule

Do not paste a client's sensitive contract into a random consumer chatbot. Not because every chatbot is malicious, but because:

• You are probably under an NDA that prohibits disclosing that information to a third party without consent.
• Free consumer products do not always behave like API endpoints with respect to data use, and those settings change every few months.
• You have no record of what went where.

If the contract contains personal information about individuals, trade secrets, or a client's financial data, this decision is not technical, it is contractual.

What okdoc does, stated plainly

When you use okdoc's AI features, such as generating a document from a prompt or analyzing a contract you received, the document content is sent to OpenAI's API for processing. That happens under OpenAI's API terms: data submitted through the API is not used to train models by default, and there is limited retention for abuse monitoring.

And here is what would not be true, so we will not say it:

• okdoc does not perform automatic redaction or masking of sensitive details before sending.
• okdoc is not configured with zero data retention with the provider.
• okdoc has no SOC 2, ISO or PCI certification.

That is the honest picture, and it is the information you need to make a decision. If you operate under a strict confidentiality undertaking, or under a regulation that requires a full subprocessor map, verify that before you run an AI feature on a client document.

Practical guardrails

• Separate drafting from data. You can ask for structure and clauses without pasting the real names and numbers, and fill those in afterward. That covers a large share of real use cases.
• Third-party personal data. Social security numbers, bank details, health information. Think twice, and consider masking.
• A document you received under NDA. Check what the NDA says about third parties and cloud services, before rather than after.
• Decide as a policy, not case by case. Decide once which categories of document may pass through AI and which may not, and write it down. Ad hoc decisions under time pressure are how things leak.

Security and permissions: what a connected assistant can and cannot do

Connecting an AI assistant to your business systems is a real change in your security model. It is worth understanding.

This is an AGENCY entitlement. API access, webhooks and the MCP connector are part of AGENCY ₪499. On FREE and CLOSE everything else works from the interface, including AI generation and the full Track suite, but without external programmatic access.

What the assistant can do. Quite a lot. It can pull an account overview, create and update documents, send them, send reminders, send from saved templates, bulk-send to a list of recipients, manage contacts, folders and templates, turn automations on and off, pull revenue reports and awaiting-payment pipelines, analyze a contract you received, and download a sealed signed copy.

What it cannot do. Sign on anyone's behalf. Reach another organization's data, because every call is scoped to the organization the key belongs to. Edit a document that has already been signed or paid. Delete a document or a contact without an explicit confirmation, because destructive operations require a dedicated confirm flag and fail without it.

Read-only scope. You can issue a key with read-only scope, and the server then blocks every tool that is not a pure read. Such a key can look at the account, the stats, the document list and search, the contacts, the templates, the automations list and the payments, and it cannot create, send, change or delete anything. That is the right choice for anything that is really reporting, and for any integration you are still evaluating.

An API key is a credential, not a setting. Whoever holds it can do anything its scope permits, in your business's name. Do not paste it into a shared chat, a Google Doc, or shell history. If it leaks, rotate it. Where the assistant supports a proper sign-in flow, prefer that over a pasted key.

Rate limiting. There is a ceiling on calls per minute, both per IP and per organization. That is a safety net against a runaway loop.

A habit worth adopting. Start with a read-only key. Let the assistant work for a week in read mode only. Watch what it actually does. Then let it write. And once you do, keep the draft-first default. The combination of "can write" and "sends by itself" is what turns a small mistake into a mistake in front of a client.

One week, in practice

So this does not stay theoretical, here is what it looks like at a small agency:

Monday morning. A sales call went well. The prompt: "Draft a proposal from this brief, in the style of the last proposal I sent, and leave it as a draft." Two minutes. Open it, fix the scope language, verify names and amounts, add a 40% deposit at signing, send over WhatsApp.

Monday evening. WhatsApp alert: the client is opening the proposal. They are reading right now. Do nothing, just know.

Tuesday. Track shows two opens, five minutes total, most of it on the pricing page, scroll depth 70%. Never reached the signature. This is not a cold deal, it is a deal with one open question about price. Short message: here is the option to split it into three payments.

Wednesday. The client asks for a small scope change. Update the document and resend. The quota is not consumed twice for a draft, only for what goes out.

Thursday. Signed, and the deposit was charged on the same screen. The sealed copy is stored, and the webhook updates the internal tracker.

Friday. Three other proposals are flagged as stalled: sent, never opened. A check finds two of them in spam. They go out again over WhatsApp.

Friday afternoon. One prompt: "Give me a weekly summary: how many sent, how many signed, how much collected, and what's awaiting payment." One minute.

Notice what happened. The AI saved time at the beginning and at the end. What closed the deal was knowing what happened in the middle.

Real limitations: when this is the wrong tool

For this guide to be worth anything, it has to say when not to:

• Complex negotiation with counsel on both sides. That process is redlines and versions, not fast drafting.
• Documents that require notarization or a legally mandated form. Some transactions require a specific format, and some categories are excluded from ESIGN and UETA coverage entirely, such as wills and certain court documents.
• Heavily regulated sectors. If a regulator dictates the language, use the regulator's language.
• An agreement you do not understand. If you read it and it is not clear to you, an AI draft did not solve the problem, it hid it behind smooth phrasing.
• One document a year. If you get a client signed once a quarter, building a workflow will not pay for itself. Use free PDF signing and move on.
• Absolute confidentiality. Documents that must not leave your systems in any form. If that is your situation, write them by hand.

What you need to get started

• FREE ₪0. 5 deals a month, AI document generation, digital signature with a sealed copy, multi-signer, OTP verification, email and WhatsApp delivery, sign and pay, templates, and the entire Track suite. No credit card required.
• CLOSE ₪199. 50 deals, automations and adaptive reminders, and landing-style signing pages.
• AGENCY ₪499. 200 deals, 10 seats, plus REST API, webhooks and the MCP connector.

The practical recommendation: start on FREE, run five real deals, and see where your process actually breaks. If it breaks at follow-up, you need automations. If it breaks on volume, you need a bigger quota. If you want to run the whole thing from inside a chat or wire it into other systems, you need AGENCY. Sign up here.

Further reading: contracts and agreements, digital signatures, okdoc inside ChatGPT and Claude, and the industry pages for lawyers, agencies, accountants and real estate.

Frequently asked questions

Can an AI assistant sign on my behalf, or on my client's behalf? No. It prepares, sends, reminds and reports. The signature is always made by the signer, on their own device, after verification. That is a deliberate design decision, not a gap that will be closed later.

Is a contract drafted by AI legally valid? A contract's validity does not depend on who typed it, it depends on the content and on the parties' agreement. A contract drafted with AI help, reviewed by you, and properly signed is a contract. That is exactly why the human review beat is part of the process rather than a suggestion.

Does this replace a lawyer? No, and anyone telling you otherwise is selling something. It shortens the time to a first draft, and it shortens the time to read an incoming contract. The legal decisions, especially on meaningful amounts and in regulated fields, are made by someone with a license.

Is my document content sent to an external AI provider? Yes, when you use an AI feature. Generating a document from a prompt and analyzing a contract both send the content to OpenAI's API for processing, under OpenAI's API terms: no training use by default, and limited retention for abuse monitoring. okdoc does not perform automatic redaction and is not configured with zero data retention. If that does not match the sensitivity of a given document, do not run AI on it and write it manually.

Does okdoc have SOC 2 or ISO certification? No. okdoc holds no SOC 2, ISO or PCI certification, and we will not claim otherwise. Documents are stored with organization-scoped access control, and every document carries a full event log.

How is this different from just asking ChatGPT to write a contract? Plain ChatGPT produces text. It does not produce a document with recipients, signature fields, a sealed copy, open tracking and payment. With okdoc, the output is a real document in your account, and you can see what happened to it after you sent it.

Will the connector send documents without my approval? Not by default. Creating a document through the connector produces a draft and returns a dashboard link. Immediate sending requires an explicit request and a recipient. The habit worth keeping: always ask for a draft, even when you are in a hurry.

What if the assistant makes a mistake and deletes something? Destructive operations require a separate explicit confirmation and fail without it, precisely to prevent this. On top of that, a read-only scoped key blocks every write operation up front.

What does each AI action cost? AI actions are metered in credits rather than gated by plan. Generating a document with AI and analyzing a contract each cost one credit. You can check your balance at any time. AI generation itself is available starting on FREE.

Does Track cost extra? No. Opens, dwell time, time per page, scroll depth, the click heatmap, deal heat and the WhatsApp alert are part of every plan, including FREE ₪0. It is not an upgrade.

Can I run this across dozens of clients at once? Yes, through bulk send from a saved template, up to 50 recipients per call. Note that each recipient counts as a deal against your monthly quota.

Which is better, an offer timer or reminders? It depends on the deal. A timer creates real time pressure and fits a proposal with a genuine expiry. Reminders fit documents that need to be signed anyway, like an agreement already verbally agreed and just not executed. Using both on the same document tends to read as aggressive.

What do I do when a client opened and read it but did not sign? That is the most valuable signal you get. It says there is one open question, not a lack of interest. Look at where they spent the time in Track and address that specific point. In these cases a short call beats an automated reminder almost every time.

Which languages does this work in? Any language your assistant speaks, and okdoc itself operates in English, Hebrew, Arabic, Russian and French, including full right-to-left documents.

Can I connect this to my CRM? On AGENCY, yes. Webhooks fire on `document.sent`, `document.viewed`, `document.signed` and `payment.received`, each signed so your endpoint can verify it came from okdoc.

Bottom line

A good AI contract workflow is not "the computer writes my agreements." It is a division of labor: the AI produces drafts and extracts information faster than any person can, you make the decisions and read the clauses somebody could sue on, and the system shows you what your client actually did with the document.

That last link is what most people miss. Speeding up the writing is easy. Not guessing about what happens after you hit send is harder, and worth considerably more.

Open a free okdoc account, run one deal end to end, and see the difference for yourself.