Generative AI consulting sells slides. The P&L needs shipped software.

Most generative AI consulting engagements hand over documents (a use-case inventory, a deck, a proof of concept on sample data) that rarely move a P&L number. Write the SOW so the deliverable is running software in one workflow. Future Works ships it in 12-week cycles against a baseline Finance signs, with a portion of the fee staked on the result.

In August 2025, Fortune reported on The GenAI Divide: State of AI in Business 2025, from MIT's NANDA initiative. Based on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public AI deployments, it found that about 5% of AI pilot programs achieve rapid revenue acceleration, while the large majority stall with little or no measurable impact on the P&L. The lead author, Aditya Challapally, put the core issue down to a "learning gap" for tools and organizations: generic tools stall in enterprise use because they do not learn from or adapt to the workflow.

What does a generative AI consulting engagement deliver?

A typical engagement delivers four things, in roughly this order: a use-case inventory, a strategy deck, a proof of concept and a proposal for phase two.

The use-case inventory

The first few weeks go to interviews and workshops across functions, and the output is a spreadsheet of ideas scored for value and feasibility. It is a fair way to see the options, and it is also where the bias starts. The values are estimates from people who will not run the workflow, and nobody in Finance has checked them. MIT found that more than half of generative AI budgets go to sales and marketing tools, while the biggest returns showed up in back-office automation. If the scores come from a workshop, check whether any back-office idea made the top of the list.

The strategy deck

Next comes the roadmap: a target operating model, a governance structure, sometimes a center of excellence, and a sequence of waves stretching across a year or two. A deck like this helps a leadership team agree on direction. It also starts a long clock, with value promised for the later waves. As I wrote in September, a long roadmap asks a board to fund the plan on faith, because it can describe value for years before anyone has to produce it.

The weak point in this phase is the handoff. The people who wrote the deck are often not the people who build, and I have written before about what happens next: the build team re-interviews the same stakeholders, rediscovers the same constraints, and some of the context is lost at every pass.

The proof of concept

Then comes a proof of concept, usually on a clean extract of data, with a few willing users, judged in a demo on accuracy. It answers a real question, whether the model can do the task. In 2026 that is rarely the question that decides the outcome. The Stanford Digital Economy Lab's Enterprise AI Playbook, published in April 2026 and built from 51 deployments that delivered measurable value, found that 77% of the hardest challenges were change management, data quality and process redesign. In 42% of the cases the model itself was fully interchangeable. A proof of concept that tests the model spends the budget on the part least likely to go wrong, and leaves integration, exception handling and adoption untested.

The last deliverable is often a proposal for phase two. Each step was billed on effort, and none of them put anything at stake on the result. The people doing this work are usually capable, and the sequence is rational for a firm paid by the hour: the contract asked for documents and a demo, and that is what arrived.

How does generative AI deliver ROI?

Software changes the P&L when it runs inside a workflow people use every day, on live data and real volume, with one named person in the business who owns it after go-live, and a number that was measured before the build started. The MIT findings point the same way. Among the factors Fortune lists for success: giving line managers, as well as central AI labs, the job of driving adoption, and choosing tools that can integrate deeply into the workflow and adapt over time. Both are about where the software lives and who owns it, which is the part a deck and a demo leave out.

How do you write a generative AI consulting SOW that ships software?

Most of this can be fixed in the statement of work, before anyone starts. These are the lines I would put in, in this order:

1. One workflow and one metric, named in the first paragraph. "Invoice exception handling time in accounts payable" is a scope. "Generative AI use cases across finance" is a study.
2. A baseline for that metric, taken from your systems and agreed in writing with your Finance team before the build starts, including how it will be measured at the end.
3. A definition of "delivered" that means running in production, on live data, with the real users, for a stated number of weeks, which a demo on sample data does not meet.
4. A named operator in your business who owns the workflow after go-live and works with the build team from the start.
5. A portion of the fee staked on the measured result, and a fixed clock of a quarter at most for the first cycle.

Expect pushback, and some of it is fair. A provider asked to stake part of its fee on one number will want access to your data early, a say in scope, and a business owner who turns up to the working sessions. Your Finance team has to give real time in the first two weeks. Revenue is often the bigger prize and also noisier inside a quarter, so agree up front how a revenue effect will be attributed.

How does Future Works run generative AI consulting?

Those five lines describe how we contract at Future Works, and they are why we call our work AI transformation with skin in the game. Your Finance team signs an Approved Value Baseline in week 2 of the 12-week cycle, the software is running in your operation from week 4, and by week 12 Finance has checked the value against that baseline. We put a portion of the fee on a value target your Finance team validates. I wrote about why large time-and-materials firms struggle to offer this.

The build is done by senior experts whose names are on the SOW, each directing AI agents on our proprietary AI-native operating system, and none of them is swapped out without your sign-off. Our pilot page shows what a first cycle ships.

A proposal you have in hand can be tested the same way. Go through its deliverables section and mark each item as a document or as running software with a number attached. Where every item is a document, send the five lines above back to the provider before anything is signed.