An AI implementation roadmap for an operating company should be a ranked list of business processes, each with a measured baseline, a date it goes into production and a person who owns the result. Many of the roadmaps companies receive are strategy documents instead, and many AI programs end in the gap between a working pilot and a process that runs every day. S&P Global Market Intelligence's 2025 survey of 1,006 IT and business professionals in North America and Europe found that 42 percent of companies had abandoned most of their AI initiatives, up from 17 percent a year earlier, and that the average organization scrapped 46 percent of its proofs of concept between proof of concept and broad adoption. A roadmap built around processes and production is designed to get through that gap.
What an AI implementation roadmap should contain
The roadmap a company typically receives after an AI strategy engagement lists use cases by department, scores them on value and feasibility, and recommends a data platform, a governance committee and a training program. Each of those can be useful, and none of them puts a process into production. A roadmap an operator can run from has four things the strategy version usually lacks. The first is a ranked list of processes, named the way the finance or operations team names them (supplier invoice processing, cash application, the bank reconciliation), with the volume and exception rate of each. The second is a baseline for the first process on that list, measured before any work starts. The third is a date for that process to run in production on live work, with real documents, real systems and a person accountable for the output. The fourth is the measured result that decides whether the second process starts.
McKinsey's survey data supports building the roadmap around processes. In its 2025 research on how organizations capture value from generative AI, McKinsey found that redesigning workflows had the biggest effect on whether organizations saw EBIT impact from generative AI, out of 25 attributes the firm tested. A use-case list describes where AI could help. A process list says which workflow changes first, how it will change, and how the company will know it worked.
How to pick the first processes
Two numbers do most of the work in ranking candidates: volume and exception rate. Volume sets the size of the return, since a process that runs thousands of times a month repays the engineering and one that runs a dozen times doesn't. The exception rate sets the cost of getting it right. Every back-office process has a clean path and a tail of cases that don't fit it, and the tail is where the hours go. In accounts payable, Ardent Partners' State of ePayables 2025 put the average invoice exception rate at 18.4 percent, so roughly one invoice in five needs a person to resolve it before it can be paid.
The best first process has high volume, an exception rate someone can measure, rules the team can already state, and a system of record where the result gets written. A high exception rate doesn't rule a process out. A process where 30 percent of cases are exceptions can still be a strong candidate when those exceptions fall into a handful of recurring types that an agent can investigate and hand to a person with the evidence attached. A process where most cases turn on judgment is a poor first candidate however painful it is, because its result will be hard to measure and harder to trust. We set out the selection tests in more detail in back office automation: which processes to automate first.
The roadmap, in order
The order of the steps matters more than the length of the list. For a mid-market company, the roadmap should run like this.
- Rank three to five candidate processes by volume and exception rate, and choose one to go first.
- Baseline that process before anything changes: hours per month, cost per transaction, exception rate and cycle time, taken from the systems and from watching the work as well as from interviews.
- Ship the first process to production on live work, with a person reviewing every case the agent can't complete with confidence.
- Measure the same four numbers after the process has run for a full cycle, usually a month or one close.
- Start the next process once the measured result holds, and carry over what the first one taught about the systems and the data.
The baseline is the step most often skipped, and it decides every argument that follows. We describe how to take one in how to baseline a process well enough to stand behind a result. Without a baseline, a program that works can't prove it, and a program that doesn't work can't be stopped early. Gartner's 2026 survey of finance leaders found that 63 percent of finance organizations said AI implementation was slower than expected in 2025, and Gartner's advice was to measure AI by the value realized instead of by the number of deployments. A roadmap built on baselines produces that measure for each process it touches.
The production date belongs in the roadmap from the first page because it changes how the work gets done. A pilot can run on a sample of documents, a copy of the data and a team that checks every output. Production means the full range of documents the business receives, writes into the live ERP, and an exception path that reaches a person who can resolve the case. Committing to the date at the start makes the team deal with integration and exceptions in the first month, while there is still time to change the design.
Why roadmaps stall between pilot and production
Generative AI implementation work tends to stop in the same place. The pilot shows that a model can read invoices, classify tickets or draft a reconciliation on a curated sample, and that result is real. What the pilot doesn't test is the distance between a high share of correct answers and a process the controller is willing to sign off. A pilot that gets 85 percent of cases right looks close to done, but the remaining 15 percent holds the unusual formats, the missing data and the conflicting records, and those cases take most of the engineering. We describe that distance in the 80-to-99% problem.
Three pieces of work usually go unfunded in that gap. The pilot read exported files, and the production process has to read from and write to the ERP, the bank portal and the shared inbox, which is integration work the pilot never scoped. Someone has to receive the cases the agent can't resolve, with the evidence attached, and that exception path has to be designed, staffed and measured like any other part of the process. And a production process needs a named person accountable for its output, plus someone responsible for keeping it working as suppliers, document formats and systems change. McKinsey's 2025 State of AI survey found that nearly two-thirds of respondents said their organizations had not yet begun scaling AI across the enterprise. A roadmap that names the integration, the exception path and the owner for each process puts those costs in the plan before the pilot starts.
What an AI implementation consultant should deliver against it
An AI implementation consultant earns the fee by moving processes from the roadmap into production, and the deliverables in the contract should be written in those terms. For each process the engagement covers, the company should receive the baseline, measured and agreed with the process owner; the process running in production on live work, connected to the systems of record; an exception path with a named person at the end of it; and a measured result after a full cycle, reported in the same four numbers as the baseline. The handover should include documentation of how the process runs and how to change it, along with the log of every action the agent took.
A consultant whose final deliverable is a roadmap and a recommendation to start a pilot has delivered the part of the work that costs least and carries least risk, which is the pattern we describe in why your AI consultants left you a deck. Three things in a proposal separate the two kinds of engagement: what will run in production, by which date, and which number will be measured against which baseline. A proposal that can't state all three is a proposal for advice.
Pricing follows from the same logic. A fee based on hours pays more for a longer engagement. A fee tied in part to the measured change against the baseline gives the consultant a reason to get the process into production and keep it there, and it gives the buyer a direct reading of how confident the consultant is in the result.
A reasonable counter, answered
A reasonable counter is that a company needs the strategy work first: a data foundation, a governance policy, an assessment of skills and a portfolio of use cases across the business, and that shipping one process ahead of all that produces an isolated result that doesn't scale. The concern is fair for a large enterprise running dozens of AI programs across business units, where uncoordinated projects duplicate spending and add risk. For a mid-market company the order runs the other way. The first process in production produces the data work that matters (the integration to the ERP, the cleaned vendor master, the list of exception types), the governance controls that matter (approval limits, an action log, a way to reverse an entry), and a measured result that tells the board what the second process is worth. A strategy written after one process has run in production rests on the company's own numbers. A strategy written before it rests on estimates.
Common questions
- What should an AI implementation roadmap include?
- It should include a ranked list of the business processes the company will change, with the monthly volume and exception rate of each. It needs a baseline for the first process, measured before any work starts: hours per month, cost per transaction, exception rate and cycle time. It needs a date for that process to run in production on live work, and a named person accountable for the result. And it should say what measured result has to hold before the second process starts.
- What does an AI implementation consultant do?
- An AI implementation consultant puts AI into a company's working processes, as opposed to advising on where it could go. The useful deliverables are a measured baseline, a process running in production against the company's real systems and documents, an exception path that routes hard cases to a named person, and a result measured against the baseline after a full cycle. A consultant whose final deliverable is a roadmap and a pilot proposal has done the planning half of the job. Ask in the proposal what will be in production, by when, and which number will be measured.
- Why do AI pilots fail to reach production?
- A pilot usually runs on a curated sample, exported files and a team that checks every output, so it never meets the integration, the full range of documents or the exception queue that production brings. S&P Global Market Intelligence's 2025 survey found that the average organization scrapped 46 percent of its AI proofs of concept between proof of concept and broad adoption. In our experience the common gaps are writing results into the live systems, a designed path for the cases the model can't resolve, and a named owner after go-live. A roadmap that commits to a production date at the start forces the team to deal with all three in the first month.