Takeaway: A widely used AI feature is not automatically profitable. A clever AI feature is not automatically defensible. The serious test is whether the workflow improves a business outcome, completes that outcome at healthy cost, and becomes harder to replace as it runs.
AI products now face two questions at the same time.
First: does this create measurable value?
Second: why can this not be copied, bought cheaper, or rebuilt internally with coding agents?
Those questions are uncomfortable because many AI features feel useful before they are economically sound. A feature can save time and still fail to move a business metric. A copilot can get heavy usage and still leak margin. An agent can look impressive in a demo and still be easy for a competitor, platform, or internal team to clone.
This is the ROI reality check.
Not ROI as a slide-deck promise. ROI as a production discipline: outcome, cost per completed workflow, pricing model, adoption path, and moat.
The Signal
The market is starting to separate AI activity from AI value.
BCG’s 2025 AI value research found a small group of “future-built” companies generating substantial value, while a much larger group reported minimal revenue and cost gains. BCG’s practical explanation matters more than the percentage: the companies capturing value are not just adding AI tools. They are reshaping how work gets done. (BCG, 2025)
McKinsey’s 2025 State of AI survey points in the same direction. Its AI high performers were much more likely to redesign workflows, track impact, embed AI into business processes, and scale agents beyond isolated experiments. (McKinsey, 2025)
Gartner has been blunter about agentic AI. It predicted that more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. Gartner also warned that many agentic projects remain early experiments or proofs of concept. (Gartner, 2025)
At the same time, coding agents are changing the buyer’s alternative. Retool’s 2026 Build vs. Buy Report found that 35% of surveyed teams had already replaced at least one SaaS tool with a custom build, and 78% expected to build more custom internal tools in 2026. (Retool, 2026) Gartner’s 2026 enterprise AI coding agent research also frames the market as moving from AI-assisted development toward agentic workflows across the software development lifecycle, with governance, pricing, support, workflows, and commercial clarity becoming part of enterprise evaluation. (Gartner, 2026)
On pricing, Bessemer’s AI monetization playbook states the hard shift directly: unlike classic SaaS, every AI query carries material cost, including compute, inference, and often human review. Pricing must account for those unit costs while capturing the value created. (Bessemer, 2026)
Put those signals together and the lesson is clear:
Usage is not enough. A better demo is not enough. Even a genuinely helpful AI feature is not enough if the economics and defensibility are weak.
The Buyer Mentality Shift
The old software buying default was simple enough: buy the platform, configure it, train the team, and absorb the license cost.
That default is weakening when coding agents can build internal tools in matter of hours or days.
Not because every company suddenly wants to become a software company. Most do not. Custom software still has maintenance cost, security work, integration debt, support burden, and governance risk. Yes, a rushed internal build can become expensive in a quieter way. But the buyer’s calculation has changed. When coding agents reduce the cost of building a narrow internal workflow, buyers start asking sharper questions: “Why buy the whole platform if we only use one part of it?”, “Why wait six months for a vendor roadmap if our team can build the missing piece?”.
This does not mean every SaaS product gets replaced by internal tools. That would be too simple. It means vendors lose the right to be vague about value.
IMHO, the message to teams building consumer-facing AI products is:
- Good: If the product owns a valuable workflow, integrates deeply into systems of record, protects data and permissions, improves with operational memory, and produces a measurable result, it still has a reason to exist.
- Weak: If the product is mostly a model call, a chat box, and a nicer interface, the buyer’s build option is strong.
Activity Is Not Value
Most internal AI dashboards measure motion: prompts run, tokens consumed, summaries generated, agents deployed, and workflows launched.
These are not useless numbers. They can help teams understand cost and operational load. But they are activity metrics and do not prove the work is more valuable.
Durable AI value connects to the business:
- support cost per ticket
- time to resolution
- conversion or retention lift
- onboarding or implementation time
- revenue per employee
- gross margin per customer
- issues or error reduction
- manual review hours removed
- faster sales cycle
Strong teams define clear business outcomes and find a way to measure them. Always take the time to carefully define the metrics. The question is not did the AI produce something? The question is what changed downstream because of it?
A useful rule:
If the AI metric never reaches a P&L line, an operating dashboard, a customer renewal conversation, or a risk review, it is probably activity dressed up as progress.
The Margin Trap
Traditional software has been built on a beautiful assumption: the next user is cheap.
Once the product (internal or external) existed, serving one more user or one more action often had tiny marginal cost. Agentic workflows break that assumption. A serious AI workflow may include:
- document retrieval
- embedding search
- reranking
- a long-context model call
- few tool calls
- structured output validation
- a second model call for critique or classification
- retries when the model output is malformed
- observability traces
- eval sampling
- human review for uncertain cases
- infrastructure, queueing, and storage
The uncomfortable part is that the true AI cost is often hidden across systems. It is not “the cost of a prompt”; it’s a cost stack.
There are also credible reports that LLM cost are significantly subsidized by the providers and that increased token usage and inference costs from agentic workflows, will result in cost increases to the end users soon.
One advice: monitor the end-to-end cost of a completed workflow…
Measure Cost Per Completed Workflow
Any AI product, internal or external, needs a workflow cost ledger.
Not a monthly API bill. Not a rough token estimate. A ledger that can reconstruct what it cost to complete one valuable job for one workflow or one customer.
For every workflow run, tag the cost-bearing events:
- input and output tokens
- retrieval calls
- tool calls
- retries
- human review minutes
- observability and eval sampling
- infrastructure time
Then attach those events to the business object the customer cares about:
- resolved tickets
- qualified leads generated
- PRs merged
Then connect that completed unit to the company metric it is supposed to influence. This is the metric chain. This means the team can explain the path from workflow cost to business value without fake precision.
The number that matters is:
What did it cost to complete the outcome, and what business value did that outcome plausibly create?
Once we know this, the decisional conversations change meaningfully.
You can route simple cases to cheaper models and escalate only when needed. You can cache repeated retrievals. You can replace model judgment with deterministic code where the rule is known. You can cap retries. You can shorten context. You can decide where human review is worth the cost. You can price in a way that matches actual usage. Without this ledger, the team usually argues from vibes — reducing the trust-factor internally which bleeds externally.
Pricing Has To Match The Cost Shape
This section applies to teams building consumer-facing AI products. AI pricing is hard because customers want predictability while vendors need protection from unbounded usage.
That is why old seat-based pricing often gets strained. Seat pricing can work when usage is predictable or the AI cost is small relative to contract value. It gets dangerous when a small group of power users can generate most of the cost.
Usage-based pricing has the opposite problem. It protects the vendor, but customers may hesitate if the bill feels unpredictable.
Outcome pricing is appealing when the outcome is clean: resolved ticket, qualified lead, completed claim, booked meeting, remediated vulnerability. It aligns value and price. But it requires confidence in attribution, quality, abuse controls, and cost variability.
For many AI products, the practical answer is a hybrid:
- base subscription for access, workflow integration, admin controls, security, and support
- included usage for predictable adoption
- usage or outcome tiers above a threshold
- enterprise negotiated terms for heavy-volumes or high-risk usage
- clear limits rather than surprise throttling
The deeper rule is simpler than the packaging:
If usage doubles and your cost doubles, the revenue model needs some way to adjust.
The AI Value Equation
Before funding, buying, or scaling an internal AI product, run the value equation.
1. Business Outcome
What number should improve?
Use a number the business already cares about: cycle time, support cost, conversion, retention, onboarding speed, implementation cost, margin, revenue per employee, error rate, incident rate, or customer satisfaction.
If the outcome is “better productivity,” define where that productivity lands. Otherwise it will evaporate into busy calendars.
2. Baseline And Target
What happens today, and what should happen after the AI workflow is adopted?
Do not skip the baseline. Without it, every improvement claim becomes mushy.
Baseline:
- current volume
- current cost
- current cycle time
- current error rate
- current staffing model
- current quality threshold
Target:
- expected improvement
- time horizon
- owner
- acceptable risk
- measurement method
3. Cost Per Completed Workflow
What does one successful outcome cost when fully loaded?
Include inference, retrieval, tool calls, retries, evals, observability, infrastructure, human review, customer support, and implementation work where relevant.
If the product needs humans in the loop, include them. Hiding labor does not improve margin. It only delays the conversation.
4. Funding Model Alignment
Does the way the workflow is funded match the way it consumes cost and creates value?
Internal products still need an economic model. If cost scales with usage but budget is fixed, a successful workflow can become politically expensive. If the bill is invisible, teams may overuse it. If usage is tightly capped, adoption may stall before the company learns where the value is.
The point is not to create a complex chargeback scheme on day one. The point is to make the mismatch visible: who benefits, who pays, what usage is reasonable, and what level of value justifies broader rollout.
5. Adoption Path
Who changes their behavior for the value to appear?
AI ROI often fails because the product works but the workflow around it does not change. The team still copies output into another system. Managers still ask for the old report. Employees still route around the tool because the official path is not trusted. Reviewers still redo the work because they do not believe the AI output.
Value does not come from the model. It comes from the changed workflow.
6. Risk And Trust Cost
What controls are required before the workflow can run?
A support triage workflow has a different cost profile than an internal drafting assistant. A financial approval workflow has a different trust bar than a meeting summary. A code migration agent has a different risk surface than a documentation assistant.
Include the cost of controls: evals, audit logs, permissions, human review, rollback, security review, incident response, and ongoing monitoring.
If the trust cost is ignored, the ROI case is incomplete.
7. Compounding Advantage
What gets stronger as the product runs?
The best answer is not “more usage.” The best answer is compounding advantage:
- better workflow data
- better evals
- better operational memory
- deeper integrations
- stronger internal trust
- more precise implementation patterns
- lower cost per outcome
- faster adoption by adjacent teams
If nothing compounds, the product is an internal shortcut with a support burden.
That may still be worth doing. But it is not a durable operating asset.
A Simple Value Gate
Use this before a team scales an internal AI product.
- Outcome: Can we name the business metric this workflow improves?
- Baseline: Do we know the current cost, time, quality, and volume?
- Unit: Have we defined the completed workflow, not just the model call?
- Cost: Do we know the fully loaded cost per successful workflow?
- Funding: Does the budget model make sense if usage increases?
- Workflow: Has the surrounding process changed enough for value to appear?
- Controls: Are trust, review, audit, and rollback costs included?
- Compounding: What gets cheaper, safer, or more valuable after every workflow run?
- Decision: Are we scaling, narrowing, funding differently, rebuilding, or killing it?
The decision rule is simple:
If you cannot answer 1 through 4, you do not have an ROI case.
If you fail 5, success may become a budget problem.
If you fail 6 or 7, the workflow may not survive production.
If you fail 8, the product will not compound.
What Strong Teams Do Differently
Strong teams do not wait for finance to discover the economics.
They instrument cost from the first production workflow. They know which model calls are necessary and which are habit. They route by difficulty. They use deterministic code where AI judgment is not needed. They cache what should not be regenerated. They cap retries. They treat human review as a designed control, not an invisible subsidy.
They also avoid the vanity of generic AI features.
When they build, they build around a process that matters. When they buy, they ask whether the vendor fits the company’s workflow, data, trust, and implementation needs. When they scale, they align funding with the value delivered and the cost incurred.
Most importantly, they are willing to narrow.
The strongest internal AI products are often not broad assistants. They are narrow workflow systems with excellent context, tight permissions, measurable outcomes, and a cost model that can survive real usage.
Narrow is not small. Narrow is how the economics become visible.
The Point
AI adoption gets serious when the conversation shifts from features to economics.
Not economics as a spreadsheet after the build is done. Economics as product design.
A durable internal AI product has a clear business outcome, a known cost per completed workflow, funding that respects the cost shape, an adoption path that changes real work, and compounding advantage from workflow ownership, governed data, operational memory, and trust.
The weak products will still get usage. Some will get impressive usage. They will be popular, expensive, fragile, and hard to defend when budgets tighten.
Activity is not value.
Usage is not margin.
Novelty is not operating leverage.
The workflow is the product. The economics are the business.
Related reading: