The pattern breaking a lot of AI budgets

Here's a story we've now seen play out in four different enterprises in the past twelve months.

Q1: Executive team decides AI agents are strategic. Budget approved. Vendors selected. Pilots launched.

Q2: Pilots demo well. Everyone is impressed.

Q3: Bills start arriving. LLM costs are higher than modeled. The productivity numbers are hard to isolate. Adoption is uneven.

Q4: CFO asks pointed questions. Nobody has good answers. Renewal decision is deferred. Or worse, killed.

McKinsey's mid-2026 research reported that a substantial share of agentic AI projects are at risk of cancellation by the end of 2027. Not because agents don't work — they do — but because the initial ROI stories weren't real, and by the time the renewal conversation came, no one could produce the number that would justify the spend.

This is a solvable problem. But it takes discipline that most projects skip in the pilot phase. Here's the framework we use to structure agentic ROI conversations — the one we walk CFOs through, and the one that keeps agent programs alive through the second-year renewal.

The four questions that decide the renewal

When a CFO evaluates renewal, they are asking four questions, whether or not they say them out loud. If your project has answers, you renew. If it doesn't, you don't.

1. What specific outcome did we buy?
2. How would we know if we got it?
3. What did it cost us — fully loaded?
4. What's the counterfactual — what would have happened without it?

Every ROI conversation should be able to answer these four questions with specific numbers. Vendors will not answer them for you. Your team has to.

Question 1: what specific outcome did we buy?

Vague outcomes ("increase productivity," "improve customer satisfaction," "modernize operations") do not survive a renewal conversation. Specific outcomes do.

Good outcome statements look like:

• "Reduce tier-one support ticket volume by 40% within six months."
• "Cut sales-research time per opportunity from 45 minutes to 10 minutes."
• "Automate invoice processing for 80% of vendors."
• "Reduce first-response time on inbound leads from 4 hours to 15 minutes."

Notice they're all measurable. They have a baseline (implicit or explicit). They have a target. They have a time bound.

The single most common mistake in agent projects is starting without a specific outcome. If you can't complete the sentence "we bought this agent to [do X] by [Y date]," you'll fail Question 1 automatically at renewal.

Question 2: how would we know if we got it?

For every outcome you claim, you need a measurement that's:

• Baseline-anchored. You must know what the number was before the agent. Ideally with historical data going back a year.
• Continuously measured. Not "we'll check at the end." Weekly or monthly dashboards that leadership sees.
• Attribution-clean. The agent's contribution has to be separable from other changes.

The attribution problem is real. If you deploy an agent at the same time you hire more support reps, run a marketing campaign, and change your pricing, you cannot isolate the agent's contribution. Do agent rollouts as clean experiments — cohort-based, A/B-tested where possible, with a stable control group.

We recommend running new agents in a "shadow mode" for two weeks before turning them on for real users. In shadow mode, the agent processes the request in parallel with the human process, but its answer is not shown. This gives you a clean baseline of what the agent would have done, versus what actually happened. Compare during and after rollout to isolate impact.

Question 3: what did it cost us — fully loaded?

Vendor pricing is one component. Fully-loaded cost includes:

• Model/API costs. Tokens consumed, per request, per month. Model your worst-case scenarios; unexpected usage spikes are the #1 cause of budget surprises.
• Infrastructure costs. If you self-host any of the stack (vector store, orchestration, observability), factor these in.
• Integration costs. The one-time engineering cost of connecting the agent to your systems. Usually 3-6 months of one senior engineer's time for a non-trivial deployment. Don't amortize this over a year — model it as a real capitalized expense.
• Ongoing engineering. Every agent needs maintenance. Bug fixes, evals, prompt updates when models change, escalation handling. Budget one engineer per five significant production agents, minimum.
• Change management. Training the humans who work with the agent. Sometimes 3-5% of affected employees' time for the first six months. This is a real cost that never appears on a vendor invoice.
• Compliance and legal review. For regulated industries, initial and ongoing legal review costs. Non-trivial for finance, healthcare, and government workloads.
• Failure costs. Some percentage of agent decisions will be wrong. Model the cost of those mistakes and any remediation.

The gap between "vendor cost" and "fully loaded cost" is typically 2-4× for enterprise agent deployments. Model the real number.

Question 4: what's the counterfactual?

This is the question CFOs care about most and vendors avoid most.

If you saved 10,000 support hours with an agent, three things could be true:

• Best case: You redeployed those 10,000 hours to higher-value work. The full labor cost of those hours is a real savings.
• Middle case: You held headcount flat and used the agent to absorb growth. Savings equal what you would have spent hiring.
• Worst case: You laid off no one, didn't grow, and just have 10,000 fewer hours of work. Savings are close to zero on a P&L basis (you're still paying the salaries), though morale and capacity have improved.

Which of these applies depends on your business. The mistake is claiming Best Case when Worst Case is what actually happened. CFOs know. They see the payroll numbers.

The disciplined approach: model the counterfactual explicitly in the ROI case. "If we did not deploy this agent, we would have needed to hire N additional headcount to handle projected growth. Agent cost is $X. Fully-loaded cost of N additional hires is $Y. Net savings: $Y - $X."

That kind of case survives a renewal review. "We saved 10,000 hours" does not.

The three ROI patterns that actually pay back

We've seen many agent deployments up close. Three patterns consistently produce clean, defensible ROI. Most others don't.

Pattern 1: Deflection

The agent handles a defined volume of work that used to require humans. Support ticket triage, invoice categorization, meeting note generation, first-response drafts. These are unglamorous but their ROI is clean because the volume is measurable and the labor cost is well-understood.

Typical payback: 6-9 months on well-designed deflection agents. Beyond that, ongoing savings compound as volume grows.

Pattern 2: Force multiplication

The agent doesn't replace anyone. It makes existing employees measurably more productive at a specific task. Sales-research agents, code-review agents, contract-review agents, marketing-brief generators. The savings come from senior people doing more senior work while the agent handles the preparatory grind.

Typical payback: 4-8 months, harder to measure than deflection but often larger in absolute terms.

Pattern 3: New capability

The agent enables something the company couldn't do before. Real-time competitor monitoring at scale, personalized outbound at every account, always-on customer research summaries. The ROI here is not cost savings but revenue generation, and it takes longer to prove — 12-18 months typically.

Most projects that fail at renewal were sold as Pattern 3 (new capability) but never produced measurable new revenue. If you're pitching a Pattern 3 agent, be honest that the ROI is a longer-cycle bet.

The two mistakes that kill ROI cases

Mistake 1: pretending soft benefits are hard ROI.

"Improved employee morale," "better decision-making," "more strategic focus" — these are real benefits but they don't survive a CFO's spreadsheet. If your ROI case leans on soft benefits, expect the renewal to be a fight. Include them as bonus, not primary.

Mistake 2: not planning for model-cost inflation and adoption tail.

Two years into a deployment, your monthly bill is roughly 2-3× what it was in month one. Why? Users get comfortable with the agent, throw more at it. Product teams expand the use cases. Model providers push you to newer, more expensive tiers. The 2026 CFO who was told "this will cost $10K/month" is now paying $28K and asking hard questions. Model the tail cost from day one.

The one-page ROI template we give clients

Every agent project we ship comes with this template completed at the outset. We update it quarterly.

Agent name: Business problem: Specific outcome + baseline + target + timeline: Measurement method (baseline data source, ongoing dashboard): Fully-loaded cost — Year 1 / Year 2 / Year 3: Counterfactual scenario (what would happen without it): Net financial impact — Year 1 / Year 2 / Year 3: Soft benefits (list, not counted in ROI number): Renewal decision criteria (what must be true to renew at end of year 1): Kill criteria (what would cause us to shut this down early):

If you cannot fill in this template today for every agent in your organization, you have exactly the setup that produces cancellation surprises at renewal time. Fix it now, not later.

Frequently asked questions

Are there standard ROI benchmarks by use case? Ranges are more honest than benchmarks. Deflection agents in mature deployments typically show 30-50% cost reduction on the target workflow. Force-multiplier agents typically show 20-40% productivity gain for the target user. New-capability agents vary wildly and shouldn't be benchmarked against averages.

How long should Year 1 be? Twelve months from full production rollout, not from pilot start. Give the agent time to reach steady state, and give your team time to change behavior around it.

Should we pilot with one vendor and roll out with another? Sometimes. Pilots are cheap to run against multiple vendors. Full deployments have real switching costs. Do vendor selection carefully — the pilot winner is often the deployment loser once integration complexity is factored.

What about compliance costs — how do we model those? Get the legal and compliance review done early, and treat it as a fixed capex, not a per-use expense. For regulated industries (finance, health, GCC government) this can be a substantial one-time cost. Amortize over the expected deployment life.

What if our organization can't measure baseline? That's Question 2 asking you to invest in measurement first. Deploy telemetry before the agent. Two weeks of clean baseline data usually costs less than a month of guessing at ROI later.

The strategic read

Agentic AI is not a science project. It's a strategic investment that requires the same rigor as any other capex decision. The teams that treat it that way — with real outcome statements, real measurement, real fully-loaded costs, and real counterfactual analysis — get the renewals and expand the programs. The teams that treat it as a "let's see what happens" investment mostly do not, and the failed programs become internal cautionary tales that make the next investment harder.

At Xenolve we help enterprise clients frame ROI for AI-agent deployments before the first line of code is written, and instrument the measurement so that Year 1 renewal conversations are grounded in numbers, not narratives. If your team is preparing an AI-agent business case for a 2027 budget cycle and wants a pressure-tested framework, get in touch — this is exactly the kind of engagement where an honest outside perspective saves everyone a lot of time.

The agents will keep coming. The question is which of them you'll be able to defend in the renewal meeting.