How do you de-risk an AI project before you commit?
Commit in stages, each ending in a decision you can make on evidence: a map of the work, a written standard, a baseline and agreed stop rules.
Commit in stages, and make every stage end in a decision you can take on evidence. Before the build, get a map of the work, a written standard on real examples and a baseline. Agree up front what result means scale, what means fix and what means stop. Start with one piece of work, not a programme. No one can honestly guarantee the impact of an AI project before they have seen the work, but you can make sure you never have to guess whether it is working.
Why the question comes up
Buyers ask it in two forms. On recent calls one asked whether the impact is guaranteed or depends on the audit, and whether there is a decision moment after the audit. Another wanted the vendor to have skin in the game, because the fear is paying for a polished slide deck and nothing else. A third wanted to move from an abstract conversation to a pilot on one process.
The fear is reasonable. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. MIT's 2025 survey found 95% of organisations see no measurable P&L impact from their generative AI pilots. Most of those projects did not fail on the model. They failed because nobody could say what the work was, what good looked like or when to stop.
Can the impact be guaranteed?
Not honestly, not before anyone has seen the work. Impact depends on things nobody knows at the first meeting: how the work really runs, which sources it depends on, how messy the exceptions are and what the team will accept. A guarantee given before that is either a guess or so narrow that it means little.
What you can ask for instead is a way to know, at every stage, whether the project is on track. That means three things in writing before the build starts:
- The standard. Real examples of the work, including the messy ones, and what a good result looks like for each. See agree the standard before you build.
- The baseline. How the work performs today on quality, cost, time and human effort, so a later result can be compared with something.
- The thresholds. What result means scale, fix or stop, agreed before anyone is attached to the build. The method is in how to measure AI agent ROI.
With those three, the question "is it working?" becomes a reading, not a debate.
Commit in stages
The cheapest way to cut risk is to never commit to more than the next decision needs.
| Stage | What you commit | What you get | The decision at the end |
|---|---|---|---|
| First conversation | A short call | Whether the work is a fit at all | Is one piece of work worth a closer look? |
| Audit | A short, fixed period with the people who do the work | A map of the work, the first build, the standard it must meet and a baseline | Build the first piece of work, change the scope, or stop |
| First build cycle | One piece of work | The first job tested on real examples against the standard | Go live, fix, or stop |
| Each later cycle | The next scope | Results against the same standard | Extend, hold, or stop |
The audit decision is the one most buyers skip. It should be a real choice, made on the map, the first scope and the baseline, not on how good the demo looked. If the audit shows that the work is not ready, because the sources are missing or nobody can say what good looks like, that is a cheap result, not a failed project. The AI readiness assessment lists the questions that decide it.
Make the first scope small and real
A pilot on one real process tells you more than any amount of strategy. Pick work that runs often, that someone owns and that has a result you can check, such as preparing account meetings or chasing open quotes. Keep it narrow enough to finish in one cycle. Gartner's advice points the same way: many use cases sold as agentic do not need agents at all, and in many cases rethinking the workflow from the ground up is the better path. That is why the redesign comes before the build: redesign the work before you automate it.
A first scope that tries to cover a whole department is the most common way to turn a small, cheap decision into a large, expensive one.
What to agree before the build starts
- Who owns the work inside the company and can make decisions.
- Which systems and data the first job needs, and that access is agreed.
- The standard, the baseline and the thresholds, in writing.
- What the agent may do on its own and what waits for a person's approval.
- When results are reviewed, and who decides at each review.
If a vendor cannot put those on paper before the build, the risk sits with you, whatever the contract says about guarantees. The buyer checklist for choosing a forward deployed engineering partner covers the rest of what to ask.
How Ortelian approaches it
Every engagement starts with the audit, about two weeks on site, which sets the scope of the first build. You leave it with a map of the work, the first build and the standard it must meet, and a baseline to measure against. Real examples of the work become evaluations, tracked on quality, cost, time and human effort. Delivery then runs in two-week cycles: every two weeks the results are reviewed together and the next step is agreed, and the first job goes live when it passes the evaluations agreed at the start. Scope, support and commercial terms are agreed before work begins. More on how we work and the platform.
Questions people ask
Can an AI vendor guarantee results?
Not honestly before it has seen the work. What it can commit to is a written standard on real examples, a baseline and a review after every cycle, so you can see whether the result is coming and stop if it is not.
Is there a go/no-go decision after an AI audit?
There should be. The audit should end with a map of the work, a first scope, a standard and a baseline, and the decision to build should be taken on those. If the audit shows the work is not ready, stopping there is the cheap outcome.
Should we start with a pilot or a proof of concept?
Start with a first piece of real work, judged against a written standard. A proof of concept that only shows the technology can work tells you little; a pilot on one process with a baseline tells you whether it pays.
How small can a first AI project be?
Small enough to finish in one build cycle: one piece of work, one owner, the few sources it needs and a result you can check against real examples.
Sources
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 2025
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, 2025
- Ortelian, How we work and Platform
