How to Rescue a Stalled AI Project or Exit It Cleanly

Pause new spending for two weeks and take stock: list what was built, where it lives, who holds each login, what you've paid and what's still missing. Rescue the project only if one small piece can be live within 30 days with a named owner. Otherwise exit cleanly: recover the accounts, export your work, switch automations off, then cancel.

The trap at this point is sunk cost. What you've already paid is gone whichever way you decide, so it shouldn't count. Compare only what finishing will cost from today, in fees, tool bills and your own hours, with what the finished piece would be worth over the next year. If a pilot stalled inside the business with no outside supplier involved, the causes are usually internal, and diagnosing why an AI pilot stalled is the better starting point.

Follow me on Instagram@sagnikteaches

Freeze the meter before you decide anything

A stalled project keeps costing money quietly: a freelancer logging hours on "tidying up", subscriptions renewing, an automation checking a spreadsheet every few minutes. Your first job is to stop anything that costs money without producing value, without switching off anything customers or staff rely on today.

Connect on LinkedInSagnik Bhattacharya

• Pause supplier work in writing. Ask for no new tasks and no new billable hours until a set date, usually two weeks out.
• Leave running anything already in daily use. If the enquiry form feeds a spreadsheet that staff work from, it stays on.
• List the subscriptions and their renewal dates. An annual renewal that falls inside the pause changes your timetable.
• Book the decision meeting now, for the day the pause ends. Pauses without a date drift into quiet abandonment.

The pause message should ask for the facts you'll need, so the two weeks aren't spent waiting. A version you can adapt:

Subscribe on YouTube@codingliquids

Subject: Pausing work on [project name] until [date]

Hi [first name],

We're pausing new work on [project name] from today until [date]
while we review where it stands. Please don't start new tasks or
bill further hours during the pause.

To help us decide how to proceed, could you send by [date]:
1. A list of everything built so far and where each part lives
   (which tool, which account).
2. Which accounts are in your name, and what's needed to move them
   to ours.
3. The state of each part: working, partly working or not started.
4. Anything you're waiting on from us.

We'll come back to you by [date] with either a revised scope to
finish a smaller version, or a plan to wrap up and hand over.

Thanks,
[your name]

Take a project inventory in one sitting

Before judging the project, write down what physically exists. Most owners discover at this step that key parts live in the supplier's accounts, paid on the supplier's card, with no documentation. That changes the risk of every option, so find out now rather than in the middle of an argument.

Here's a filled-in inventory for an illustrative tutoring agency: the owner, two coordinators, about 60 tutors on the books and around 180 active students. Six months earlier the agency hired a freelancer to automate tutor matching, so that each parent enquiry would produce a list of available tutors and a drafted introduction email. Two of the three milestone payments had been made when progress stopped.

Two things jump out. The part that failed, matching, was failing because of the availability data rather than the AI: tutors weren't updating the sheet, so no system could know who was free. And the part that worked, drafting, sat inside someone else's account. Both findings shape the decision more than any argument about the freelancer's pace.

Where stalled AI projects usually jam

Name the jam precisely, because each one has a different cost to clear. Projects with an outside supplier tend to stall at one of these six points:

1. Waiting on your side. The supplier needs data, access or decisions that nobody inside the business has time to provide. The cost to clear it is your hours, not theirs.
2. An accuracy plateau. The demo looked good, but on real inputs it's right perhaps seven times in ten, and every fix breaks something else. This is often a data problem wearing an AI costume, as it was for the tutoring agency.
3. Scope that grew. "While you're in there, could it also…" turned a four-week job into a vague one with no finish line.
4. Supplier capacity. Replies slow from days to fortnights because a bigger client arrived. It's common with solo freelancers and rarely fixed by chasing.
5. The ground moved. The tool or feature the project was built on has changed or is being retired. OpenAI's custom GPTs stop running on 11 December 2026, for example, so a project built on one has a deadline whether it stalled or not.
6. Budget or attention ran out. The owner's priorities shifted and nobody inside picked the project up.

Jams 1 to 3 are usually rescuable once they're named. Jam 4 means changing who does the work. Jams 5 and 6 are where a clean exit is most often the right answer.

The rescue-or-exit test, with the sums

Ask four questions. Rescue only if all four answers are yes; otherwise plan the exit.

1. Is there one piece that could be live and in daily use within 30 days?
2. Is there a named person inside the business with about two hours a week to own it?
3. Is the cost of finishing that piece, counted from today, less than a year of the value it would produce?
4. Can someone, the original supplier or a replacement, commit to the date in writing?

Question 3 is where the sums come in. At the tutoring agency, coordinators spent about 25 minutes per new enquiry checking availability and writing the introduction, at roughly 30 enquiries a month: about 12.5 hours a month. Suppose the finished drafting piece brings that down to 8 minutes including the coordinator's check. That saves about 8.5 hours a month, or roughly 100 hours a year. At an illustrative $25 an hour for coordinator time, the piece is worth about $2,500 a year.

Finishing it from today would take perhaps 15 internal hours (about $375 at the same rate), a year of the automation tool (Make's paid plans start from about $9 a month, so call it $110 a year) and whatever the supplier's remaining work costs. Any remaining supplier fee comfortably under about $2,000 therefore still leaves the rescue worth doing in the first year. The two milestone payments already made appear nowhere in that sum, and they shouldn't.

The same sum run on the full original scope, with automatic matching, failed question 1: it couldn't be live in 30 days because the availability data would take a whole term to fix. So the agency rescued part of the project and parked the rest, which is the most common good outcome.

Rescuing it: cut back to the smallest piece that can go live

A rescue is a new, smaller project with its own scope, date and acceptance test, not a continuation of the old one. Write the new scope down before talking to anyone, then compare it with the original. Here are the tutoring agency's before and after:

Original scope: "Automate tutor matching and parent introductions for all enquiries, check tutors' calendars, and send WhatsApp alerts to the chosen tutor."

Rescue scope: "When a coordinator enters the chosen tutor in the enquiry sheet, draft the introduction email to the parent in our template, for the coordinator to check and send. Live by [date]. Accepted when 18 of 20 test drafts need no more than a one-line edit."

The rescue scope drops the part blocked by bad data, keeps a person making the matching decision, and ends with a test anyone can run. Writing deliverables and acceptance criteria covers the wording in more depth; the principle is one trigger, one output and one pass mark.

Then reset the working arrangement:

• Price the remainder as a fixed piece of work rather than open hours, so the supplier's incentive is to finish.
• Move the working parts into accounts you own as part of the rescue, not at the end. For the agency, that meant its own Make account and its own API key, with the freelancer invited in as a user.
• Hold a 15-minute check every week until go-live: what was done, what's blocked and whether the date still holds.
• Agree in advance what happens if the date slips by more than a week. Usually that means stopping and switching to the exit plan.

If the original supplier can't commit, a new one can pick the work up, but expect them to charge for a short review first and possibly to rebuild parts they can't vouch for. Give them what a departing supplier should have left you anyway; the AI consultant handover checklist lists it item by item.

Exiting cleanly: the shutdown order that avoids surprises

A clean exit means you leave with your data and your logins, and with nothing still running in the background. Order matters: recover access first, while the relationship is still cordial, and cancel last.

1. Recover admin access to every account. Anything created under your business email or paid with your card should be recoverable through the vendor's normal account process. For anything in the supplier's name, ask for ownership to be transferred or the work exported.
2. Export the work itself. In Make, "Export blueprint" in the scenario menu downloads a JSON file holding the modules, their settings and mapped values; connections aren't included, so they need setting up again wherever you import it (Make's help page on blueprints). Zapier can copy Zaps to another account, but connections must be re-authenticated, copies arrive switched off, run history doesn't come across and webhook URLs change (Zapier's help page on copying assets). n8n downloads a workflow as a JSON file that includes credential names, so strip those out before you share it.
3. Save the prompts and instructions as plain text. These are often the most valuable part of a failed project and the easiest to lose. Paste them into a document the business owns.
4. Get chat and project data out before closing any AI workspace. ChatGPT Business has no self-service export; OpenAI's help centre tells members to ask the workspace owner about organisation-managed data access. On Claude Team, the Primary Owner can export the organisation's data, including members' chats. Check what your plan allows while the account is still open.
5. Switch off triggers before anything else, so nothing new starts while you dismantle the rest.
6. Revoke connections and API keys, and change any passwords you shared with the supplier.
7. Cancel subscriptions in renewal-date order, checking notice periods and annual commitments first.
8. Tell staff what goes back to manual, and from when. If anything customer-facing is being removed, such as a website chat widget, update the page that pointed to it.

The automation nobody switched off

Skipping step 5 is the classic failure. An illustrative two-shop pet store abandoned a stock-alert automation in the spring but never deleted it. The Make scenario kept checking an inventory sheet every 15 minutes, and each scheduled check uses a credit even when nothing arrives. That one idle scenario used roughly 2,880 credits a month, well past the 1,000 a month that Make's free plan allowed as of September 2026, so the shop was paying for a plan to run something nobody wanted.

Zapier behaves differently but has the same blind spot: it auto-pauses a Zap only when 95% of its runs error over seven days. A Zap that runs "successfully" on stale data will carry on indefinitely. The only reliable check is a person opening each automation tool and confirming the list of active workflows is empty, or holds only the ones you meant to keep.

When the tool is being retired underneath you

Sometimes the exit is forced. An illustrative picture framer built a quoting assistant as a custom GPT, loaded with moulding prices and a sizing guide. It stalled when prices changed and nobody updated the knowledge file. Custom GPTs stop running on 11 December 2026, and OpenAI's migration route converts a GPT into a plugin, with its instructions becoming a skill and its knowledge files becoming reference files.

The framer's cheaper exit was to copy the instructions and the current price list into a ChatGPT Project, which staff can share on a business plan, and to keep a simple spreadsheet calculator as the fallback for quotes. The wider lesson, set out in keeping your data and prompts portable, is to hold every instruction and reference file outside the tool as well as inside it, so a retirement notice costs you an afternoon rather than a rebuild.

Settling up with the supplier without a fight

Most stalled projects end with an awkward conversation about the final invoice. Keep it factual and tied to the contract by answering three questions first:

• What was each payment for? Match milestones to deliverables and note which were accepted.
• Who owns what was built? Check the ownership clause; who owns the AI workflows a consultant builds explains the usual positions.
• What does ending the contract require? Look for notice periods, payment for work in progress and any handover duties.

Then propose a settlement in writing. A calm version:

Hi [first name],

Thanks for the status list. We've decided to [finish a smaller
version / wrap the project up]. Based on the contract:

- Milestones 1 and 2 (form set-up, drafting step): delivered,
  accepted and paid.
- Milestone 3 (matching): not delivered in a usable state.

We propose [paying for the handover work at the agreed rate /
closing the project with no further invoices], in exchange for:
- the Make scenarios exported as blueprints,
- the prompt text as a document,
- moving the API key's billing to our account, or deleting the key.

Could we agree this by [date]?

[your name]

If the amounts are significant or the supplier disputes the facts, ask a solicitor before sending anything that reads as formal notice. What you need most are the files and the access, and a small concession on the last invoice is usually cheaper than a dispute.

Checking, 30 days later, that the decision held

Rescues and exits both fail quietly, so put a 30-minute check in the diary a month after the decision.

• After a rescue: is the piece in daily use? Pull 20 recent outputs and count how many needed more than a one-line edit. If it's more than the pass mark allowed, you're back at jam 2 and should decide again rather than drift.
• After an exit: read the next bank or card statement line by line for any AI or automation charge you didn't expect. Open each automation tool and confirm nothing is switched on. Check the API usage page shows zero activity, and that the supplier no longer appears as a user anywhere.
• Either way: ask the two people closest to the work whether anything they relied on quietly disappeared. Exits sometimes remove a small helpful step, such as an auto-filled spreadsheet column, that nobody listed.

A one-page post-mortem that makes the next attempt cheaper

Whether you rescued or exited, write one page within a week while memories are fresh. Keep it free of blame; it exists for the next project, not as a verdict on the last one. The tutoring agency's, filled in:

Project: tutor matching and parent introductions
Dates: started March, paused September
Goal: cut enquiry handling from 25 minutes to under 10, and reply
  to parents the same day.
What got built: enquiry sheet (working), drafting step (working),
  matching (failed on availability data).
Where it jammed: tutors don't update their availability; nobody
  checked that data before the project started.
What we kept: the drafting step, now in our own Make account.
What we'd do differently: test the data on 20 real enquiries
  before signing; keep every account in our name from day one;
  hold a 15-minute check each week.
Revisit matching: next term, once tutors confirm availability
  weekly through the form.

Two lines in that note, testing the data first and holding the accounts from day one, would have prevented most of the stall. That's typical. The fixes are rarely technical, and they're cheapest to apply before the next project starts rather than halfway through it.

Rescuing or shelving a stalled AI project: follow-up questions

Can I get money back on an AI project that stalled?

It depends on the contract. Match each payment to the deliverable it was tied to: if a milestone was paid but its deliverable was never accepted, you have grounds to discuss a partial refund or a credit against finishing work. Put your summary in writing, stick to dates and deliverables rather than blame, and offer a settlement first. If the sums are significant or the facts are disputed, ask a solicitor before sending anything formal.

What if the supplier won't hand over logins or files?

Start with accounts created under your business email or paid with your card: the vendor's own account-recovery process usually gets you back in without the supplier. For accounts in the supplier's name, ask in writing for a transfer or an export and quote the ownership clause in your contract. If they still refuse, contact the tool vendor's support team, then a solicitor. Change any shared passwords the same day.

Is it ever right to finish the whole original scope?

Yes, when the jam was entirely on your side and has since been cleared: the data now exists, the decision has been made, or the person with time has been found. Even then, finish it in stages, with the first stage live within 30 days, so you find out quickly whether the rest is still worth building. Re-run the forward-cost sum for each stage rather than for the whole thing.

How do I tell staff that the project is being shelved?

Be specific and brief: say which parts are switching off, from which date, what goes back to the old way of working and which parts are staying. Thank the people who tested it and ask what they would keep. Staff who spent time on a stalled project usually care most about not losing the bits that helped them, so name those and keep them where you can.

Further reads

• Scope Creep in AI Projects: How to Handle Change Requests — Stop the rescued version growing back into the project that stalled.
• Why AI Projects Fail in Small Businesses (It's Rarely the Tech) — The organisational causes behind most stalls, and their early signs.
• How to Run a Pre-Mortem Before You Launch an AI Project — Run one before the restart so the same jam doesn't recur.
• What If Your AI Vendor Shuts Down? Checks Before You Commit — Protect yourself when the tool, not the supplier, disappears.
• How to Judge Whether Your AI Consultant Delivered Value — Work out what the work done so far was actually worth.
• Automation Audit: Find the Zaps and Scenarios Nobody Owns — Find every leftover Zap and scenario nobody owns.
• How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — Score the rescued version properly after a quarter of use.
• AI Screening: Bolt It Onto Your ATS or Switch Systems? — Bolt-on, upgrade or switch: how a small agency decides on AI screening, with switching costs worked through and a 30-day trial log to copy.
• Why AI Adoption Stalls in Accounting Firms, and How to Restart It — The five reasons AI stalls in accounting practices, a twenty-minute diagnosis, and a 30-day restart timed around the deadline calendar.
• Do Small Medical Practices Need an AI Consultant? — When a small medical practice can handle AI itself, the signs outside help will pay for itself, what a consultant shouldn't do, and a scoped brief to adapt.
• Should You Hire an Automation Consultant or Build It Yourself? — A five-question test, what DIY really costs in hours, one build that worked and one that didn't, and the hybrid route most small firms should take.
• How Long Does AI Implementation Take for a Small Business? — Realistic planning ranges for small-business AI projects, where the weeks really go, and a week-by-week timeline for a sports equipment shop.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Make help centre (scenario blueprints; credits used by scheduled triggers); Zapier help centre (copying assets between accounts; auto-pause rules); n8n documentation (exporting and importing workflows); OpenAI help centre (exporting ChatGPT data; custom GPT retirement and migration FAQ); Anthropic Claude Team plan documentation.