AI Transformation for Internal Operations: Running It as a Workflow Portfolio, Not a Programme
Most mid-size companies start their AI transformation the same way. They buy licences, run a workshop, name a task force, and six months later nobody can point to a single process that works differently. One practitioner summed up the pattern as "change no actual workflows" followed by "ask why there's no ROI." The fix is a smaller and stricter unit of progress: a portfolio of redesigned internal workflows, each with an owner, a baseline, and a retired manual path.
This guide is for COOs, operations leaders, and AI program owners at companies of roughly 50 to 1,000 people who were handed the mandate without a Big Four budget. It shows how to count progress, how many workflows your people can carry, which ones to build first, and when to kill one.
The short answer
Treat AI transformation as portfolio management. Keep one register of candidate and live workflows. Count a workflow only when the old way of doing it is gone and someone owns the new one. Limit how many live workflows each owner carries, because every one of them needs ongoing care. Build the first workflow to establish shared controls, then pick the next two for how much of that foundation they reuse. Review the register every two weeks and make explicit keep, kill, and retire decisions.
| Programme habit | Portfolio replacement |
|---|---|
| Count licences, active users, or pilots | Count workflows whose manual path is retired, with a named owner |
| Roadmap of 20 use cases ranked by ROI | Register of five to eight, sequenced for control reuse |
| Steering committee status slides | Fortnightly review with keep, kill, or retire decisions |
| Budget for build, not for running | Run cost and owner hours on every register row |
Why the enterprise playbooks do not fit a mid-size company
The pages that rank for "AI transformation" are useful, but they are written for a different organisation. IBM's explainer defines the term as integrating AI into operations, products, and services, and lists planning stages from information gathering to building a roadmap. EY's seven-layer blueprint runs from systems of record through infrastructure, knowledge, trust, process, workforce, and customer. SysArt's 2026 playbook describes five maturity levels, a centre of excellence that later federates, and a dedicated transformation team with platform, governance, and change leads.
A 200-person company does not have a platform team or a governance lead. It has an operations director, two or three people who understand the systems, and perhaps an outside builder. If you copy the enterprise structure, you spend the first quarter writing frameworks. If you skip structure, you get scattered experiments that nobody maintains.
The core advice still carries over. EY's advice is to start with one process or workflow aligned with top-down objectives, and to reuse components as you redesign the next ones. SysArt warns that most organisations overestimate their maturity by one or two stages and that pilots stall when integration and governance arrive late. The portfolio approach keeps those lessons and drops the org chart.
Define the counted unit before anything else
Count a workflow as transformed only when its old manual path is retired and a named owner carries it, not when licences are issued or a pilot demos well. The ranking pages measure adoption metrics, active users, and pilot success criteria instead, and those numbers can grow for months while the work itself stays the same.
As long as the spreadsheet, the inbox rule, or the copy-paste routine still exists, people fall back to it whenever the new workflow is slow or uncertain. You then pay for both paths, and your "transformed" count is fiction. A retired manual path forces the questions you would otherwise defer: who handles exceptions, what happens when the model is wrong, and who gets paged.
The UK Government's AI Playbook makes the ownership half of this explicit. It asks teams to nominate a senior responsible owner accountable for the use of AI in each project. AWS makes a similar point in its Five V's framework: AI becomes a capability, rather than a project, once it is integrated into daily operations with clear ownership.
Write the counted-unit definition down and get your sponsor to agree to it. A workable version reads: "A workflow counts when the previous manual method has been switched off or formally deprecated, exceptions route to a named queue, and one person is accountable for its quality and cost." Report that count to leadership instead of seat usage. One honest test, borrowed from a practitioner post quoting Kai-Fu Lee, is whether the programme has moved a number leadership already tracks.
Keep one portfolio register
Skip the transformation office and keep one register that everyone can read. Five to eight rows is plenty for the first two quarters. Each row is a workflow, not a tool or a department.
- workflow: supplier invoice intake and coding
owner: finance operations lead
status: live # candidate | building | live | retiring | killed
baseline: 6.5 min per invoice, 3 percent rework (measured over 4 weeks)
manual_path: shared inbox plus spreadsheet; retire after 30 clean days
controls_used: [service account, audit log, exception queue, eval set]
run_cost: model spend plus 3 owner hours per week (logged, not guessed)
kill_date: 2027-01-15 # review date if the baseline has not moved
next_review: 2026-10-28
Three fields do most of the work. The baseline lets you prove change; Fire in Belly's guide to measuring AI workflow ROI covers how to set one per workflow. The manual path field is where the counted unit lives. The kill date is a commitment made before you are emotionally invested in the build. Anthropic's enterprise guide recommends agreeing success metrics before launching any pilot; a kill date is the other half of that agreement.
Leave out anything you will not update. A register with fifteen columns becomes a quarterly exercise, and a quarterly register cannot steer a fortnightly cadence.
Cap the portfolio by owner run capacity
Cap the live portfolio by owner run capacity, because every shipped workflow becomes a permanent maintenance liability that a transformation plan rarely budgets. The ranking pages talk about scaling. Almost none of them talk about the ceiling on how much a small team can keep running.
The UK playbook is direct about the cost. It notes that ongoing investment is needed to maintain AI systems, update them with new data, and keep performance stable, and it expects teams to support live running, iterate, and eventually retire the service. AWS likewise counts total cost of ownership through ongoing operations and calls for dedicated teams for optimisation and maintenance.
In an internal workflow, that ongoing work is concrete. Someone reviews the exception queue. Someone notices when an upstream API changes shape. Someone handles the vendor notice that a model is being withdrawn, which our model deprecation runbook shows is a recurring event, not a one-off. Someone answers the "why did it do that?" message from a colleague.
The numbers below are a planning example to replace with your own logs. Log the real owner hours on your first two live workflows for a month. Then set a cap per owner from that number. If each live workflow takes an owner about three hours a week and the owner can spare six, that person carries two workflows, and the third waits or gets a new owner. The arithmetic is trivial. Writing it down is what stops a programme from launching ten workflows that two people cannot keep healthy.
Put run_cost and owner hours on every register row, refuse to move a row to building if its owner is already at the cap, and treat retiring a weak workflow as a way to free capacity for a stronger one.
Sequence the first three workflows for reuse
Choose your second and third workflows for the controls they can reuse from the first, not for the next-highest standalone ROI. This is our recommendation, synthesised from the sources below rather than quoted from any of them. EY notes that redesigning individual processes builds a library of reusable components that compounds value, and SysArt warns that platform decisions made for a single pilot create technical debt that compounds rapidly.
Your first live workflow forces you to solve a set of problems once: how the workflow authenticates to systems, where actions are logged, how exceptions reach a human, how outputs are evaluated before release, and who approves risky actions. Those are the same controls our pilot-to-production checklist and human exception queue guide cover in depth.
If workflow two shares none of them, perhaps because it touches a different system with a different identity model and a different review team, you pay the foundation cost again. If it reuses most of them, it ships faster, and the controls get tested under a second load before they become standard.
A practical selection rule:
- Pick workflow one for value and for a contained blast radius. It should be painful enough that people want it gone and safe enough that a wrong output is caught before it leaves the building.
- List the controls workflow one needed.
- Score each remaining candidate on how many of those controls it reuses, then break ties with expected value.
- Pick workflow three the same way, but allow one new control so the foundation grows deliberately.
Value still matters, and a workflow with high reuse and no business pull is busywork. Reuse is the tie-breaker that stops ROI ranking alone from scattering your first quarter across four unrelated foundations.
Run a fortnightly portfolio review
The review is short and has one output: a decision per row. A useful agenda takes 30 to 45 minutes.
- Live rows: has the baseline moved, is the manual path retired yet, and is the owner under the cap? Decide keep, fix, or retire.
- Building rows: is the kill date still credible? Decide continue or kill.
- Candidates: does any owner have spare capacity? If not, nothing new starts.
Killing a row should be normal. A practitioner discussing transformation notes that three of five founders interviewed were struggling with shared context and workflows that never stuck, which reads as an adoption and skills problem rather than a tooling one. A clean kill with a written reason is more useful than a pilot that quietly keeps running and draws attention away from the rows that are working.
Failure modes that look like progress
- Rising seat usage with a flat counted unit. People use the chat tool more, and no manual path has been retired. Record that as enablement and keep the counted unit at zero.
- A long roadmap and an empty "live" column. The register has 20 candidates and zero retirements after a quarter. Cut it to five and start one.
- Owners in name only. The owner field lists a director who never sees the exception queue. Reassign to the person who will triage it.
- Pilots that never hit a kill date. If the date passes without a decision, the review is not working.
- Every workflow on a different foundation. Each build has its own credentials, logs, and review process. Reuse broke down at step two.
Where outside help earns its fee
You can run this portfolio internally. Outside help is worth paying for in three places. The first is building the initial register and baselines quickly, which is what an AI audit is for: a fixed-scope pass that maps workflows and returns a costed shortlist. The second is building workflow one with the shared controls done properly, through AI workflow automation or similar. The third is the build, buy, or partner decision for each row, which our build vs buy guide walks through.
Be wary of anyone who sells a multi-year programme before workflow one is live. SysArt's own playbook says a readiness assessment should take four to eight weeks, not six months, and that an assessment taking longer has become a consulting exercise. For a mid-size company, a shorter scope is usually right.
How to verify the transformation is real
At the end of each quarter, answer these from the register, not from memory:
- How many workflows meet the counted-unit definition, and which manual paths were switched off?
- For each live workflow, has its baseline metric moved, and by how much?
- Is every owner at or under their cap, with logged hours to show it?
- Which controls did workflows two and three reuse from workflow one?
- Which rows were killed or retired, and what capacity did that free?
If you cannot answer the first question with a number and a list, stop adding candidate rows until you can.
What to do next
This week, write your counted-unit definition and get your sponsor to sign off on it. Then open a register with no more than five candidate rows, fill in owner, baseline, manual path, and kill date for each, and choose workflow one. Book the first fortnightly review for two weeks out. If you want a fixed-scope start on the register and baselines, Fire in Belly's AI audit is designed for that first pass.