Cut AI spend without a quality cliff

Who this is for

This playbook is for owner-operated companies that can see AI on the bill — bundled copilots, meeting notetakers, image tools, API usage — and cannot tie the line to a job. It is also for the operator who was told to “cut 40 to 85 percent” by a routing vendor and wants a first pass that does not wreck the few workflows that already work.

If the seats were never used, start at when AI seats do not equal AI usage. Reclaim is cheaper than routing. If the waste is two products doing the same non-AI job, read overlapping SaaS. If nobody ranked the work before the purchase, rank AI work that pays so you do not optimize a bill for a job you should stop doing.

Symptoms

If several of these are true in the same month, this is a live operational leak — not a tooling preference.

• The invoice has AI line items and nobody can name the workflow each one supports.
• Two copilots overlap and both renewals are approaching.
• A vendor slide promised a wide savings range if you add a gateway.
• Someone switched a live workflow to a cheaper model and quality fell with no way to see it except angry Slack.
• API usage spiked because a prompt loop retries in production with no cap.
• Finance asked for a cut and the only idea in the room was “turn it all off.”
• A routing layer was purchased before the company had a single eval set.

What done looks like

Done is a working operating change, not a purchased seat or a dashboard nobody opens.

• Every AI cost is tagged to a named workflow or marked unassigned and queued to cut.
• Cold seats and duplicate copilots have been reclaimed or cancelled first.
• Any workflow that will change model or route has a small eval bar on real examples.
• A cheaper tier is adopted only after the bar still passes.
• Retry loops, unbounded context, and orphaned batch jobs have an owner and a cap.
• A gateway exists only if volume and eval load already justified the extra layer.

Cost per workflow, not cost per magic

AI spend becomes unmanageable when it is discussed as a single blob. Split it. A meeting notetaker is a workflow or it is wallpaper. A copilot seat is a draft path or it is inventory. An API bill is a named job with a volume story or it is a loop you have not found. Write the job next to the cost. If you cannot, the first cut is the unassigned line — not a clever router.

Do not invent a return percentage to justify keeping a line. If the workflow is real, describe it. If the workflow is not real, cut it. A slide that says the company will save a dramatic band of spend is a sales artifact. You will not find a responsible version of “40 to 85 percent” in this playbook because those figures are not a method. They are a headline.

Cost per workflow also keeps you honest about quality. A cheaper model that destroys the one draft path the team actually uses is not a savings. It is a relocation of cost into rework and distrust.

Waste first: seats and duplicates

Before you tune models, remove what nobody uses. Cold seats are the cleanest cut. Duplicate copilots are the next. This is ordinary SaaS consolidation with an AI label. Map the job. Keep one tool. Cancel the other. Reclaim licenses that never met a “used” definition — the definition lives in the adoption playbook.

Waste also hides in settings. Uncapped retries, transcripts stored forever in a tool you do not need, and batch jobs that embed the same folder every night will move a usage bill without moving a business number. Put an owner on those knobs. A cap is a control. It is not a strategy by itself, but it stops the silent leak while you decide.

Do the reclaim before the architecture conversation. Teams skip to routers because routers feel like engineering. Cancelling a cold annual feels like administration. Administration is where most of the money is sitting.

Then tier or route — behind an eval bar

If a workflow is live and the bill is still too large after waste is gone, you may change how the work is done: a smaller model, a shorter context, a cheaper tier for first passes, a stronger model only when the first pass fails a check. That is tiering. Routing is the same idea with software in the middle.

Neither is safe without a bar. Write ten to twenty real examples the workflow already produced: a good draft, a borderline draft, a failure you caught. Decide what “still good enough” means in words a reviewer already uses. Run the cheaper path on those examples before you point production at it. If you cannot tell whether quality dropped, you are not ready to save money on that path.

The bar can be human. A reviewer scoring the same five proposals this Friday is an eval. A spreadsheet is enough. You do not need a platform to start. You need examples the business already recognizes and a person who will say “this is worse.”

Quality cliff means the cheaper path looks fine in a demo and fails on the messy cases you actually have: the angry customer, the incomplete note, the exception. Keep those cases in the bar. If the cheap path fails them, keep the expensive path for those cases or keep the expensive path entirely. Partial routing is a later refinement, not a day-one requirement.

A worked cut, then a careful tier

A firm has a bundled copilot on every seat, a second writing tool three people like, and an API bill from a prototype nobody turned off. Start with the map. The prototype is unassigned: cap it or shut it down. The second writing tool is a duplicate if the bundled copilot already drafts. Reclaim the cold bundled seats using the adoption rule — used means a draft entered a real path. None of that needs a new model.

Suppose one workflow remains: a first-pass proposal draft that a principal already reviews. The API or the premium tier for that path is the visible cost. Write ten real past drafts the principal accepted or rewrote. Run a cheaper model on those ten in a side-by-side. If the principal says the cheap path misses the offer structure on the messy ones, keep the current path. If the cheap path holds, switch only that workflow and watch the next five live drafts. That is a tier. It is not a company-wide model swap and it is not a savings headline.

If someone then pitches a gateway to capture a wide savings band, ask which workflows have evals and which seats are still cold. If the honest answer is “we will find out after we install it,” you are being sold a layer. Finish the waste pass first.

A gateway is a later layer

A routing gateway — one hop that picks a model, logs spend, and applies policies — can be the right build when many workflows share volume and you already have evals you trust. It is the wrong first purchase when you have twelve cold seats and one API key with no owner.

Gateways add failure modes. They can go down. They can send the wrong route. They can hide spend in a new invoice line that finance does not recognize. Buy one when the volume of calls and the cost of running evals by hand are both obviously large. If you have to invent the volume story, you do not have it yet.

Ranking still comes first. Rank AI work that pays before you invest in infrastructure to make a low-value workflow cheaper. Some bills should go to zero because the job should go to zero.

What people usually get wrong

They lead with a savings range. Wide percentage bands are how vendors make a gateway feel inevitable. Ask what they counted. Ask what quality they measured. If the answer is a slide, keep your wallet closed.

They switch the production model on a Friday because a benchmark looked good. Public benchmarks are not your proposals, your inbox, or your tone. Your examples are.

They cut the one workflow the team actually adopted because it was visible on the bill. Visible and wasteful are different. Cut unassigned and duplicate first.

They treat “turn it off” as the only control. Off is correct for unused inventory. Off is reckless for a path with a reviewer and a fallback that the business now depends on. Tier that path. Do not vandalize it.

DIY vs hire

DIY the waste pass this month. Export the AI lines from the last two statements. Tag each to a workflow or to “unassigned.” Reclaim cold seats. Cancel one duplicate. Put a cap on any retry loop you can find. That work does not need a gateway.

DIY a tiny eval only if a live workflow must get cheaper. Ten real examples. One reviewer. One cheaper path in a side-by-side. Keep or revert based on what the reviewer says, not based on a token price.

Hire help when spend spans several vendors and nobody owns the map, when a gateway purchase is already on the calendar, or when a cheap-model switch already damaged a customer-facing path. The $997 assessment is a live consultation and a written analysis. You leave with two to four improvements you can run yourself — a cost map, a reclaim list, an eval sketch — and a later roadmap quoted only if you want us on routing or a harder rebuild. Implementation is not in the fee.

Controls before automation

Recommendations that touch customers, money, contracts, private data, or systems of record need human review, limited permissions, a log, and a fallback. Do not automate a broken path because the tool is ready.

• Do not change a production model or route without an eval bar on real examples.
• Do not auto-send cheaper drafts to customers to “see what happens.”
• Cap retries, batch jobs, and context size; unbounded loops are not an experiment.
• Do not buy a gateway until waste is reclaimed and at least one eval exists.
• Keep a fallback to the prior model or to a human-only path.
• If two tools still charge for the same job, stop adding routes. Cut the duplicate.

How this sits next to the other playbooks

Cost reduction is the other half of the waste room in the fourteen mid-market AI asks. Adoption tells you which seats never became jobs. SaaS consolidation tells you which logos are duplicates. Ranking tells you whether the remaining spend is attached to work that should exist at all. None of those guides will promise a savings band. They will help you stop paying for fog.

If you want the map written with you before the next renewal or the next gateway demo, book the assessment. Bring the invoices and one workflow you must not break. Leave with a cut list and a bar.