Distill: experiments in using fewer AI tokens
An experiment showing why token savings need to be assessed alongside the quality of the resulting work. It does not yet demonstrate financial savings.
All the ways to work together · Versione italiana
From bill to decision
AI Cost Audit
Reconstruct where AI spend goes, what useful work it produces, and what is worth testing to reduce it without hurting service quality.
€10,000 excl. VAT · 2 weeks
A good fit when
Finance, product or engineering teams with one or two live AI applications, significant or unusual spend, and an economic decision to make.
The problem it solves
Bills are rising, but it is hard to see which applications, models and repeated attempts generate the spend, or what it costs to produce a usable result.
What you are actually buying
A picture of spend checked against bills and usage data. Where the data allows, we also calculate cost per accepted result: work that meets the agreed quality criteria. You receive recommendations on what to test, what to leave alone and which data is missing.
Included activities
- Reconciled cost baseline with explained variances
- Allocation by provider, model, workload and team
- Cost per accepted outcome, or an explicit measurement gap
- Options ranked by expected savings, confidence, effort and risk
What I need from you
- Billing and usage exports, preferably without live credentials
- A minimum map of accounts, workloads and owners
- Retry, failure and accepted-outcome logs where available
- Quality, security and continuity constraints that must not degrade
How it works
- Fix the scope, period and decision
- Reconcile cost, usage and allocation
- Measure waste and outcome cost without filling evidence gaps
- Compare options and choose one test with a stop rule
How the work is verified
- Reconciled total or an explained variance
- Repeatable formulas, period, currency and allocations
- Retries and rejected outputs excluded from accepted value
- Every option states estimated savings, confidence, effort and risk
Who does the work
Fifteen years building software people rely on.
- 15+: years in software engineering
- Seed → IPO: product leadership through a public listing
- 3.2★ → 4.7★: App Store rating during the team's work at bunq
- SDA Bocconi: two sessions on applied and agentic AI
These results describe my experience. For your project, we define goals and measures around your company’s process.
Portfolio and roles · All services compared, with prices and next step
Scope limits
- Up to 2 providers, 2 workloads and 3 stakeholders
- Up to 90 days of already exportable data
- One primary accepted-outcome definition
- A third provider/workload, data remediation and implementation are quoted separately
What we do not promise
- No guarantee of savings, ROI or payback
- Not an accounting audit, certification or procurement engagement
- Implementation, vendor negotiation and continuous monitoring are out of scope
- Missing data remains unavailable: it is not estimated to fill the report
Essential terms
Fixed price · Remote · up to 2 providers and 2 AI applications or processes
50% at start · 50% at delivery · proposal valid for 14 days
Cloud, tools, travel and data remediation excluded
One clarification call within 14 days of delivery
Fit assessment within 2 business days
When the work is accepted
- Costs reconciled or residual variance made explicit
- Repeatable allocation and formulas with visible missing data
- Options comparable by estimated savings, confidence, effort and risk
- One immediate decision with test, success, stop and rollback
Related evidence
How to reduce AI costs without hurting service quality
Where to start cutting AI costs: unnecessary calls, model choice, caching and context, with worked examples and a 14-day comparison.
B2B service. A request creates no commitment: price and scope become binding only in an accepted written proposal.