Three systems I shipped this year went from a Monday decision to running in production inside six weeks. Two others, pitched with far bigger budgets, still are not live eighteen months on. The gap was never the model or the vendor. It was the order everything happened in.
How do you implement AI in business without joining the pile of stalled pilots? Pick one process that already carries a number, chased invoices, answered calls, qualified leads. Audit it, check whether the data and the people behind it can actually carry a build, then run the system in shadow mode next to whoever does the job today before it acts alone. Get that order right and the technology rarely is the problem. Get it wrong and no model rescues you.
Everything below is the sequence itself, not the theory behind it: the six phases in order, the table I would hand a client on day one, why the failures nearly always show up before a line of code is written, and the one budget line almost nobody sets aside.
Six phases, in this order, every time. The publishing engine that produces this article runs on the same sequence we sold to clients, and the WhatsApp deal manager built for an estate agency (the CRM piece covers that build in full) followed it just as strictly. Skipping a phase to save a week is how the eighteen-month projects happen.
1. Audit. Map the real process, not the org chart's version of it. An AI audit exists precisely to stop you funding the wrong use case before anything else starts.
2. Readiness check. Confirm the data behind that process actually exists, is reachable, and is clean enough to trust. The readiness assessment and the shorter readiness checklist both do this job; use whichever fits the hour you have.
3. Pick one process. Name the job in a single sentence with a number already attached. Not "improve customer service". "Cut the time to answer a call from four rings to one."
4. Shadow mode. Run the system next to the person doing the job today. Nothing goes live. You are comparing two answers to the same question, honestly, for as long as it takes to trust the pattern.
5. Cutover. Let the system act, with a defined checkpoint where a person still reviews anything above a threshold you set in advance, not after the first mistake.
6. Scale. Only once the first process has held its number for a month does a second one start. Not before, however tempting the second idea looks.
Notice what is missing from that list: choosing a model, comparing vendors, picking an architecture. Those decisions matter, but they sit inside phase four, not before phase one. Reverse the order, start with the technology choice, and you end up defending a platform you bought before you knew which process it was for.
| Phase | What actually happens | Typical time |
|---|---|---|
| Audit | Map the process as it really runs and rank where automation would pay | A few days to two weeks |
| Readiness check | Confirm the data exists, is reachable, and someone can own the result | A few days |
| Pick one process | Name the job in one sentence, with a measurable number attached | Half a day, done properly |
| Shadow mode | Run the system alongside the current process without switching anything over | Two to four weeks |
| Cutover | The system acts; a person still checks anything above an agreed threshold | Ongoing, reviewed monthly |
| Scale | Start the second process once the first has held its number for a month | Repeat the cycle |
Large IT projects run over budget by 45% over budget and deliver 56% less value than predicted, according to research by McKinsey and the University of Oxford, and unclear requirements sit behind roughly two in five failures. Neither number describes a model reasoning badly. Both describe a project that started with an ambition instead of a sentence.
Agentic projects fare no better once the scope grows past one process. Over 40% of agentic AI projects will be cancelled by the end of 2027, Gartner forecasts, and 88% of AI proofs-of-concept never reach widescale deployment. I have killed two projects for exactly this reason: a goal described as a mission statement, and data that lived in three systems that never spoke to each other. Neither failure showed up in a demo. Both showed up three months in, once shadow mode exposed what the pitch had glossed over.
Ask most businesses to plan an AI rollout and the plan is entirely technical: which model, which vendor, which integration. That is the wrong half of the work. In a survey of 1,107 professionals, 38% of implementation difficulty comes from user proficiency, against about 16% from technical issues. Training, trust and a named owner beat the choice of model by more than two to one.
The readiness numbers back this up from a different angle. 42% say AI is capable but their organisation is not set up to capture the value, and 22% name the way their organisation operates as the single biggest barrier to AI success. Buy the best model available and hand it to a team with no training and no named owner, and the technology will sit there, unused, exactly as capable and exactly as useless as before you bought it.
The audit and the readiness check are not paperwork to get through before the "real" work starts. They are the real work, done first, because a build on the wrong process or on data nobody can reach fails regardless of how well it is coded. If you already know precisely which process to pick and the data is clean, skip straight to shadow mode. Most businesses do not know that yet, which is why the diagnosis earns its place. The maturity assessment is worth a look too, if only to confirm which rung of the ladder you are actually standing on before committing budget.
Treat the audit result as a shortlist, not a single verdict. A good one ranks several candidate processes by value and feasibility, and the sequence above then runs once per item, starting with whichever sits at the top. Running two items through the sequence at once, to save time, is how a business ends up with two half-finished pilots instead of one that actually works.
Shadow mode is the phase clients most want to cut, and the one I refuse to cut. It costs a few weeks and produces nothing visible, which makes it an easy target for a budget review. What it buys instead is a comparison you can trust before a single customer notices anything changed. Our own approach to the checkpoint that follows it is covered in the piece on human oversight that does not become a bottleneck, and the groundwork that makes shadow mode possible at all, breaking a job into steps a system can actually be measured against, is covered in the note on process mapping for a digital workforce.
None of this needs a large team behind it. What a normal week running agentic systems across a small group of companies actually looks like is in a normal Tuesday, and the six-phase sequence above is exactly what runs underneath it.
Audit first to find where automation would actually pay, check readiness to confirm the data and people can carry it, pick one process with a number attached, run it in shadow mode next to the person doing the job today, then cut over and only add a second process once the first holds its number.
A single, well-scoped process usually reaches production in six to ten weeks: a few days for the audit and readiness check, a few weeks in shadow mode, then a cutover. Multi-process, group-wide rollouts take months, not weeks, and should be treated as several small projects, not one large one.
Rarely the model. Vague goals, data that cannot be reached cleanly, and nobody named to own the output account for most of it, which is why 95% of generative AI pilots deliver zero measurable return. The failure gets discovered late because the order was skipped, not because the technology broke.
If you already have one obvious, well-understood process with clean data behind it, you can move straight to a small build. An audit earns its place when several candidates compete for budget or the data picture is unclear, because it turns a guess into a ranked, costed shortlist.
More than most budgets allow for. User proficiency, training and trust account for roughly 38% of implementation difficulty against about 16% for technical issues, so plan the person who owns the system and the training they need before you plan the build itself.
Pick the one process in your business with a number already attached to it and write that sentence down today: not "get better at AI", the actual job and the actual figure it should move. Everything in the sequence above exists to protect that single sentence from turning into a company-wide programme before it has proved itself once.
Bring the one process you would automate first and we will run the audit and readiness check against it live, then tell you honestly how many weeks the sequence should take.
Book a 30-minute fit call