Your AI Agency Overpromised. Here's How to Recover.
The run-up was exciting: impressive presentations, confident timelines, references that checked out. Then came the reality — missed milestones, demos that only work on the presenter's laptop, invoices that arrive more reliably than features, and lately, communication that takes days. If you're in this situation, two things are true at once: it's one of the most common situations in the current AI market, where demand for 'AI agencies' vastly exceeds the supply of teams that can actually deliver; and it is recoverable, if you move methodically instead of emotionally. This page is the method: secure your assets, get an independent verdict, then decide from facts.
Step 1: Secure your assets before anything else
Before any confrontation, quiet or otherwise, make sure you control what you'd need to walk away. One: the source code — full repository access with history, not ZIP snapshots and not a deployed black box. Two: your data and all configurations — prompts, fine-tuning setups, knowledge bases, integration settings. Three: every credential — API keys, cloud accounts, admin panels — and here's the critical part: in accounts registered to your company, not the agency's. Four: whatever documentation exists. While you do this, find your contract and read the IP clause. In a properly written contract you own the work product you've paid for. In far too many AI projects the IP situation is ambiguous — which is expensive to fix later and free to check now. If access is refused or delayed at this stage, that refusal is itself crucial diagnostic information about what's happening on the other side.
Step 2: Get an independent technical assessment
You cannot make the next decision — finish, fix, or walk away — on the basis of status reports from the people who created the situation. What you need is an independent review that answers four factual questions: Does what exists actually work? Is it architecturally sound or fundamentally flawed? How complete is it, honestly, as a percentage? And what would finishing or fixing it realistically cost? A competent independent assessment takes one to two weeks and costs in the region of €2,000–5,000. Two cautions. First, independent means independent: beware an 'assessment' from a party angling to sell you a rebuild — the incentive corrupts the verdict. Second, get findings in writing, in business language, with the technical detail attached. The document matters as much as the verdict: it's what you present to partners, boards, or in the worst case, lawyers. It's cheap insurance on a five-figure decision.
Step 3: Decide — finish, fix, or walk away
With facts in hand the decision usually clarifies quickly. The trap to avoid is sunk-cost reasoning: money already spent is gone from every option, so compare only future cost against future value. A €40,000 spend doesn't make a bad project worth €20,000 more.
| Situation per the assessment | Right move | Why |
|---|---|---|
| 70%+ sound, you own the assets | Finish with a new team | Cheapest path to value |
| Core works, quality/integration poor | Fix (targeted rescue) | Foundation has value |
| Foundation flawed or assets not yours | Walk away, rebuild clean | Second time is cheaper and faster |
| Use case itself was wrong | Stop entirely | Redirect budget to a real problem |
How to never be in this position again
Every failed engagement we've reviewed showed warning signs that contract structure would have caught. The patterns that predict failure: payments decoupled from demonstrable progress; no acceptance criteria ('working AI' is not a criterion — 'correctly processes 95% of this test set' is); code you cannot see until the end; and confidence instead of cadence — 'trust us' instead of a demo every two weeks. The corrective structure, which we would recommend for any AI engagement with anyone including us: payments tied to milestone demos of working software, repository access from day one, measurable acceptance criteria in the contract, IP assignment in unambiguous writing, and a break clause at each milestone. An agency that resists this structure has told you something important before you've spent anything. An agency that proposes it is probably safe to work with.
Related pages
- AI project failing
- How to choose AI agency coming soon
- AI consulting process coming soon
- Case studies
Frequently asked questions
The agency won't hand over the code. What are my options?+
Check your contract's IP clause first — often you're entitled to it. If genuinely unclear, legal advice is warranted, but also price the alternative: a clean rebuild is sometimes cheaper than a fight.
How do I know if what they built is actually any good?+
You can't from the inside — get an independent technical assessment. One to two weeks, written findings, factual basis for the next decision.
Can another team finish a half-built AI project?+
Usually yes, if you own the code and data. Completion is often cheaper than restart — but only an honest assessment can tell you whether the foundation deserves it.
What must an AI project contract contain?+
IP assignment, milestone payments tied to working demos, measurable acceptance criteria, continuous code access, full credential handover, and milestone break clauses.