The short answer
A realistic agentic AI pilot budget allocates roughly 15% to discovery and architecture, 35% to build and integration, 20% to evaluation and governance, 10% to change management and training, and 20% to post-launch operations through the first 90 days. Pilots that under-fund any single line stall in the same predictable ways — under-fund discovery and you scope wrong, under-fund evaluation and you cannot ship to production, under-fund change management and the team bypasses the agent.
The allocation that ships pilots
Across our pilot portfolio, the budget allocation that consistently produces shipped agents looks like this — and the budgets that stall consistently under-fund one of the same three lines.
- ›Discovery and architecture (15%): Workflow Triage, system mapping, data audit, success-criteria document.
- ›Build and integration (35%): agent logic, retrieval, tool integrations, UI, deployment.
- ›Evaluation and governance (20%): eval harness, regression suite, audit logging, human-approval design.
- ›Change management and training (10%): user training, feedback channels, documentation, op-team handoff.
- ›Post-launch operations (20%): first 90 days of monitoring, tuning, on-call, and quality sampling.
The under-funded line that stalls the most pilots
Evaluation and governance. Teams excited about the build are reluctant to spend 20% on something that does not produce visible output. The result is an agent that 'seems to work' but cannot be defended in a security review or a quality regression. Pilots that fund evaluation properly ship; pilots that do not stall in evaluation purgatory.
The hidden cost of under-funding discovery
A two-week discovery that costs 15% of the budget routinely changes which workflow gets built. Skipping discovery to save time consistently produces 4–6 week scope corrections mid-build, costing more than the discovery would have. Discovery is the highest-ROI line in the budget.
Why post-launch operations is 20%, not 5%
First-90-days operations is where the agent actually earns its eval data, surfaces real edge cases, and either earns trust with the operating team or loses it. Budgets that drop to a maintenance retainer at launch day consistently see quality drift in months 2–3 because nobody is paid to watch.
Typical pilot ranges (2026)
A small-business single-workflow pilot: $8K–$18K. A mid-market multi-system pilot: $25K–$60K. An enterprise multi-agent implementation: $80K–$200K. These ranges assume the allocation above; pilots that under-fund discovery or evaluation routinely come in at the bottom of the range and fail, or at 2× the range and recover.
Key takeaways
- 15/35/20/10/20 allocation ships pilots; deviation in either direction stalls them.
- Under-funded evaluation is the single most common cause of pilot stall.
- Discovery is the highest-ROI line — never skip it to save time.
- Budget 90 days of post-launch operations from day one, not as a separate line item later.
Frequently asked questions
Can we run a fixed-price pilot?+
Yes — most well-scoped first pilots fit a fixed price after a paid discovery. Fixed price before discovery is a red flag from any vendor; they are pricing risk into the number without understanding the actual scope.
What's the cheapest credible pilot?+
About $8K for a single-workflow small-business pilot with a tightly scoped integration. Below that, you are paying for someone to wire up an off-the-shelf tool — buy that yourself.
Should ongoing model and infra cost be in the pilot budget?+
First 90 days yes, ongoing no — it should move to an opex line owned by the operating team. Bundling them indefinitely obscures unit economics and discourages cost-aware design.