Summary: On 2026-10-15, Cohere published a guide to measuring AI ROI, written by product marketing manager Ariana Milligan. Its core argument: the math is simple, but the inputs — true costs and attributable returns — are exactly the parts AI makes fuzzy. For builders pitching or scaling AI features, the useful part is the list of places estimates quietly go wrong.
The formula is easy. The inputs are not.
Cohere defines AI ROI the standard way: (financial benefit minus total cost) divided by total cost. The difficulty is deciding what belongs in those numbers. AI costs accumulate across infrastructure and operations, while returns are tangled inside wider business processes, so isolating either takes deliberate work.
One trap the guide names: operational improvement is not the same as financial value. Faster task completion or higher-quality output only counts if that value exceeds what it cost to get. A team can ship something impressive and still lose money on it.
Unit economics that move under you
In generative and agentic systems, the cost of a task varies with the work — longer context, more reasoning, more tool calls, retries. Two runs of the “same” task can cost different amounts. Scale changes things too: production workloads are bigger and more varied than pilots, so pilot-stage unit costs rarely survive contact with production.
This is the same effect I covered in LLM tokenization pricing: identical-looking workloads can carry very different bills depending on model choice. Cohere’s suggestion, cost-aware model routing — sending each task to the cheapest model that meets its quality bar — is the practical response. Deployment choice matters just as much: usage-based managed services suit spiky demand, while dedicated capacity wins on unit economics only when workloads are high and steady.
Costs hiding outside the AI line item
The most useful section for anyone writing a business case: inference costs are only part of the real spend. Production AI also needs monitoring, automated validation, human review of uncertain cases, and — for agents — evaluation of entire action sequences, not just final outputs. High-stakes workflows need more of all of it.
Many of these costs don’t show up as AI expenses. If engineers spend part of their week maintaining integrations, or AI workloads consume centrally funded infrastructure, Cohere argues the ROI calculation should reflect those resources anyway. If you skip this, your ROI number is flattering you for costs someone else’s budget is absorbing.
When task efficiency doesn’t move the business
A task-level speedup can be real and still worth nothing. If a downstream step stays constrained, the gain just shifts the bottleneck. Cohere’s fix for attribution is methodological: use a pre-AI baseline, but don’t trust simple before-and-after comparisons — demand could have risen anyway. A/B tests, staggered rollouts, or comparing similar teams with and without AI do the isolating work.
Some gains value cleanly (more conversions, more gross profit). Others — say, a lower manufacturing defect rate — need you to trace the savings through rework, returns, and warranty costs before they mean anything.
Stress-test before you commit
The guide’s forecasting advice is refreshingly unromantic: vary assumptions about usage volume, workload mix, and value, and see which ones break the business case. Model ROI across several time horizons, since AI carries upfront costs before benefits arrive — but longer forecasts get less reliable as model capabilities and prices shift. Once you’re in production, replace assumptions with actual usage and cost data.
The practical takeaway for builders: treat your ROI model as a living document with visible assumptions, not a one-time pitch deck. And keep a suspicious eye on any forecast where inference is the only cost on the sheet.