Trust amortisation
Certify once. Trust every answer after.
AI can generate endless answers. Provenance gives your institution something more valuable: an analytical language it can shape, govern and trust.
Ask anything about Customers & accounts
Enter to send · Shift + Enter for a new line
01 / 05 · Ask
Ask in plain words. It reads the question back first. The AI named a measure itself: churn rate, counted on accounts.
Your people ask in plain words and AI interprets the question. Your institution controls the meaning. A deterministic compiler constructs new analyses from approved computational contracts. Every valid new analysis builds on trust already earned: trust compounds.
Who it’s for
For the meeting after the demo: the regulator’s letter, the audit sample, the figure someone has to sign.
- Energy & utilities
- Pharma & life sciences
- Banking & insurance
Available on AWS Marketplace
01The problem
An analyst builds it. A manager checks it. Finance checks it again. Next year an auditor samples it. AI made answers abundant, not accountable, and when the next question arrives the checking starts over. Generating another answer isn’t the breakthrough. Making the next answer cheaper to trust is.
01
Spreadsheets are handwork.
Every sheet is built by hand, so every sheet must be checked by hand. The mistakes never live in SUM. They live in the ranges and the references.
02
AI chat improvises.
Ask a new question and many AI tools write new code for it. To trust that answer, someone has to check it, then the next one, then the one after.
03
So trust never compounds.
When every report is checked on its own, you pay for confidence on every report, forever. Little of what you checked last month makes this month cheaper.
02The idea
What if the unit of assurance were not the report, but the analytical language reports are built from? Nobody inspects a lift before every ride. It is certified, then trusted. Provenance separates business meaning from computational machinery. Your institution approves what its words mean, such as “churn” at your bank. Every operation runs under a contract that says what it does and how it may be combined. That is trust amortisation: a new answer needs a quick check of what was understood, not a full review of how it was built.
1,000
Checked report by report
1,000 checks
No rules approved: every report is checked by hand
and 976 more
Review hours
3,000
Approved once
1,000 quick checks
The rules, approved once · v2026.11
and 976 more
Review hours
373
Review hours, added up
- Checked report by report
- Approved once
| Reports | Checked by hand | Approved once |
|---|---|---|
| 1 | 3 hours | 40 hours |
| 10 | 30 hours | 43 hours |
| 16 | 48 hours | 45 hours |
| 100 | 300 hours | 73 hours |
| 1,000 | 3,000 hours | 373 hours |
Illustrative: about 3 hours to check one new report by hand; about 40 hours to approve the rules once, then about 20 minutes per new answer, which also covers approving the occasional new definition. Your numbers will differ, and a pilot measures them.
Approved per version, not forever. The operations and their contracts are approved once and reused by every answer. New business meanings still need a person: each new company word is approved once, when it first appears. When a rule changes, only the change is checked.
What we tell the Chief Risk Officer
You are not being asked to trust what an AI invents tomorrow. You are being asked to trust the rules your institution approved yesterday.
03Compounding
The analytical language is already there. Your institution makes it its own.
Provenance begins with a domain-independent foundation of analytical operations, relationships and contracts. Your institution does not have to invent it. It shapes it around how its business works: what counts as churn, which customers qualify, how exposure is measured, what gets left out. These are not prompts asking an AI to remember your preferences. They become approved definitions in your institution’s analytical system, with who kept them and when.
The vocabulary grows through use. When a question reveals a missing definition, the AI proposes it. Your people challenge it, edit it and decide whether it belongs. Once kept, it is available to every analysis after.
And the system compounds.
A new definition can build on kept ones, and new questions combine approved meanings and operations in ways nobody had to specify beforehand. Only what is new needs approving, so what your institution can compute grows with every word it keeps.
You are not teaching an AI to guess how your institution works. You are shaping an analytical system that knows which rules it must obey.
The AI proposes a measure.
churn rateisaccounts closeddivided byaccounts at start
Drafted from the question and your data’s columns: “Which segments are we losing fastest?”
Your people challenge it, in plain words.
Change it in words
Customers, not accounts. Count each customer once. Leave out staff. Lower is better. Draftchurn rateisaccounts closedcustomers lostdivided byaccounts at startcustomers at start
customers lostiscountcount once per customerwhereevery account closed
only rows wherestaff=no Good when lower
The AI redrafts the meaning and the steps. Nothing changes until someone keeps it.
Someone accountable keeps it.
Every answer that names it computes it exactly as kept here.Keep this wordKept by Customer Office · 12 Nov 2026
It becomes the company’s: its meaning and its maths.
churn rate
Good when lowerMade ofcustomers lostdivided bycustomers at start
Countedonce per customeronly rows wherestaff = no
New questions, and new words, build on it.
- Which branches lose the most customers, last 12 months?churn rate
- Churn rate by month for premium customers, 2026churn rate
retention ratemade ofchurn rateBuilt on it: a new word from a kept oneComputed exactly as kept. No AI decides it again.
Drawn from the product
04How it works
Provenance never asks a language model to write SQL or a new analytical program for each question. The model proposes intent. It does not own the computational logic.
- 01
Ask in plain words.
Whatever your team would ask an analyst. No query language, and no dashboard to build first.
- 02
The AI writes down what you meant.
Your question becomes a typed specification: what to measure, over which rows, per what, in what order. It is read back in one plain sentence before any number, and you can change any part of it by hand.
- 03
Your institution governs.
The specification can only use the meanings, operations and limits your institution approved. Anything outside them is refused in plain words.
- 04
The compiler constructs.
A deterministic compiler resolves the valid operations, builds the plan and checks it against their contracts before any data moves. Code computes the result: the AI does not calculate it or write the executable analysis. Change the specification and the analysis replays, without asking the AI to invent it again.
1 · The question · in plain words
“Which regions grew fastest last quarter?”
2 · The specification · typed, editable
- Measuring
- sales growth, Q3 against Q2 2026
- Only rows where
- quarter is Q2 or Q3 2026
- One row per
- region
- Ranked by
- growth, highest first
Read back: sales growth by region, Q3 2026 against Q2 2026, highest first.
3 · The compiled plan · checked before it runs
- 1Read sales48,210 rows
- 2Keep Q2 and Q3 202648,210 → 8,904 rows
- 3Total sales by region and quarter8,904 → 24 rows
- 4Growth from Q2 to Q324 → 12 rows
- 5Sort, highest first12 → 12 rows
4 · The result · computed by code
12 regions, ranked highest first · none left out
05The audit trail
A log tells you what happened. This record also shows why the computation was allowed to happen. Before anything runs, the plan is checked against its contracts; as it runs, its steps, settings, decisions and lineage are written down, never reconstructed after. Your institution does not have to take the model’s word for it. It can inspect the machinery.
Every step, in order.
The chain the answer was made from, read left to right. Nothing happens off the record.
The exact settings.
What each step was told, exactly as the code received it. Not a summary written after.
Proven before anything runs.
The plan is checked against its structural claims before a single row moves.
Who decided what, and why.
Every choice made while building the answer is written down, with the rule that made it.
A memory that doesn’t retire.
Each step is timed and fingerprinted, so the answer can be re-made and compared exactly. People leave. The trail stays.
Provenance Held
9 steps · 2.41s
fetch_databoolean_indexboolean_indexassign_maskedgroupby_agg:standardgroupby_agg:standardmerge:siblingsassign:ratiorank
Request
Parameters
Quality
Held38 structural claims proven before any data moved.
Decisions
- COMPILER‘churn_rate’ reads customers_lost and customers_at_start per segment; each is read once per answer rowcomparison_reads_its_own_copy
- LEDGERproved 38 structural claims over 9 steps before any data movedLEDGER_LAW
Pipeline
9 steps- 0.38s
fetch_data
Data retrieval
customer_idaccount_idsegmentopenedclosedstaff
# fae1716d90644046
- 0.02s
boolean_index
Keep rows
staff = false
# 9c2e04b1aa7f3310
- + 7 more steps
As it appears under every answer
06What follows
None of these is a feature we bolted on. Each one follows from the same choice: approve the rules, not the reports.
01
Choose the model. Keep the intelligence.
Your enterprise alpha is not a set of prompts or a model’s memory. It is the meanings your people agree on, the rules they approve and the analysis they build over time. Provenance keeps it in an enterprise-governed system, apart from the AI that interprets questions.
Today the interpreter is Claude, through Amazon Bedrock. Nothing you approve lives inside it: change the model, and your definitions, decisions and analytical history stay as they are. A new model may read a question differently, and you see that in the read-back. It cannot redefine a calculation you approved.
Models come and go. Your institution’s knowledge stays.
The interpreter · replaceable
Claude, through Amazon Bedrock · today
interprets, computes nothing
Your institution’s knowledge · kept
churn rateretention rateactive customer
- Who kept each word, and when
- Saved recipes, re-made without the AI
- The trail of every answer
02
Same recipe, same answer.
Every answer is kept as a recipe. Re-run it in March or in December and get the same numbers from the same data, without the AI.
03
AI cost follows questions, not rows.
The AI reads the question and a description of your data, and code does the counting, so model cost follows the questions you ask, not the rows you hold. Ten rows or ten million.
07What we promise
No software is always right, and anyone who tells you otherwise is selling something. Auditors promise reasonable assurance, not absolute. So do we. The difference is where it is earned: on the rules, once per version, not on every report.
You own the meaning.
The read-back says what was asked, in your words. You confirm it in two seconds, or change it by hand.
We answer for the maths.
Provenance enforces the computational contracts. Every number matches the read-back, made by code that is approved once and shared by every customer.
Check us without us.
Want more assurance? Give the read-back to your own analyst, or to another AI, and compute the same thing on your own database. Same recipe, independent kitchen.
If a rule is ever wrong, it is wrong in one place. It is fixed there, and any saved answer can be re-made under the fixed rule.
08Where it runs
What a technology buyer asks first, answered with what is true today.
- Where it runs
- On AWS, in the EU (Ireland). Available on AWS Marketplace.
- The AI
- Claude, through Amazon Bedrock in the EU. It is the only setup offered today; other models, including privately hosted ones, are not. Bedrock does not use your data to train models.
- What the AI sees
- Your question and the chat; a description of your data (table and column names, types, sample values, and the values of category columns, up to 500 per column, such as segment or region names); and, when it words a report, the facts of the finished answer. It never writes code and never computes a number: every figure in its wording is filled in by the code.
- Who checks the maths
- Your own team, without us: any answer can be recomputed from its read-back on your own database.
- Independent review
- Passed the AWS Foundational Technical Review, AWS’s check of a partner product against its security and operational best practices.
What exactly do we certify once?
The rule book: what your company’s words mean, which operations exist, and the contracts that say how they may be combined. Not the reports. When a rule changes, only the change is reviewed, and every answer after is built under the new version.
Do we have to replace our dashboards or our warehouse?
No. Your dashboards, warehouse and workflows keep running. Provenance starts beside them, with the questions your team answers by hand today, and becomes the governed layer for the analyses it supports. Which of your data sources it reads first is agreed in the first meeting.
Is this a chatbot on top of our database?
No. The AI turns your question into a specification and stops there. A deterministic compiler builds the plan from approved operations and checks it before any data moves. What runs is the same approved code every time.
What if the AI misunderstands the question?
You see it in the read-back, before you rely on any number. Change the step by hand, or ask again. The read-back is printed by code from the plan itself, so a misunderstanding cannot hide behind fluent words.
Is our data used to train a model?
No. Provenance does not train models on your data. What it learns, your company words, is kept in your company’s book, where your people can read it, change it or remove it.
What happens when a question cannot be answered?
It tells you, in plain words, what is missing and why. An honest no is part of the product, not a failure of it.
A first meeting
Bring one question your team answers routinely, one unusual variation, and one that needs a new combination of analytical operations. We will show what the AI understood, what the compiler constructed, which rules were enforced, what needs human approval and what can be reused next time.
Don’t just test whether Provenance gets an answer. Test what your institution gets to keep.