A consulting deck
Two quarters of interviews that end with a document, built from what people remember doing.
How the question gets answered.
- Ontology
- Closed environments
- Creative ideation
- Data visualization
- Collaboration
- Planning
- Productivity suite
If you founded the company today, with AI, would you build this again? ZeroBase asks that of one function at a time, with the evidence attached and a named person deciding.
Start a Clean Sheet Review →“Embrace ablation: regularly delete old system prompts and scaffolding; modern models are intelligent enough to not require legacy ‘hobbling.’”
Cherny is talking about prompts. The same is true of an operation: most functions run on scaffolding written for a slower world: reports, approvals and intervals nobody would build again. ZeroBase is how you find that scaffolding and take it out with the evidence attached, before you automate it into permanence.
01 · The problem
Your board wants AI in operations. Your auditors want it to hold up.
Every function carries work designed once and inherited since: reports nobody reads, controls that never reject, intervals set at commissioning. Most of it was a workaround for one constraint: information was expensive to gather, to move and to check.
The monthly pack gathered it. The status meeting moved it. The approval chain checked it. With AI inside the systems, none of the three costs what it did, and work that only existed to pay them has lost its reason. Automating it now is the fastest way to make it permanent.
There is a fourth option.
DIY agents and scripts
Useful until it invents a figure, averages a contradiction into a score, or leaves with the person who built it.
Doing nothing
The pack keeps going to three readers, the approval keeps never rejecting, the interval stays untested.
One question, asked of a document, a meeting and a flow.
Each was built for a world where information was expensive to gather, to move or to check. The figures are from the synthetic Northstar scenario every example on this site runs on.
Document
Monthly performance packRecipients41 · 3 readersAnnual cost€186,000Copies kept4
The figures lived in seven systems and someone had to assemble them, so they were assembled once a month for everyone, whether anyone asked or not.
A question, asked when needed
The figures are queryable. One canonical copy is produced on request, and an alert goes out when a threshold moves. The pack, its four copies and its cycle are gone.
Meeting
Weekly status meetingInvited14 · 8 presentDecisions2 a quarterCadenceweekly, since 2021
Status lived in people’s heads. The only way to move it was to put the people in one room, so the room was booked every week.
A brief, and a call on exceptions
Status lives in the systems. A Monday brief is written from them, and a meeting is convened only when a decision is pending. The standing slot is gone.
Flow
Requisition approval chainSteps3 for €50Rejections below €5k0 in 18 monthsCapacity1,940 h a year
The rule needed human eyes to apply it, so every requisition went past three pairs of them, including the ones the rule could never reject.
The rule stays, the queue goes
The rule is explicit and an agent applies it. A person sees the exceptions and the values that matter. The step that never rejected anything is gone.
02 · What it is
An evidence-first operating model review.
An ontology, not a chatbot
Activities, obligations, evidence, gates and decisions are typed objects with named relations. A figure is traceable to the system, query and period that produced it, and a contradiction stays visible instead of being averaged into a score.
Bounded agents, experts in the loop
Each agent carries a may and a may-not list; anything unlisted is denied. Consequential recommendations stop at a named reviewer, screened for conflict, who can move the decision in six directions, not just approve or reject.
Runs inside your boundary
Read-only connections, metadata only, aggregated above the individual. Cloud, private cloud, on-premise or isolated: the review runs where your data is allowed to be, including environments with no outbound access.
Eight weeks, one function, four steps.
The same four steps run on every function. Each one narrows what the next is allowed to do, and nothing crosses a boundary implicitly.
01 · Connect · weeks 1 to 2
Lock the baseline, connect the sources
Pick one function and lock its baseline: annual operating cost, capacity, activity list, service levels and the systems in scope. Approved sources are connected read-only, metadata only. Nothing is measured against a moving target.
Microsoft 365
02 · Establish · weeks 2 to 4
Build the inventory from evidence
Every recurring activity gets a cost, a capacity draw, a demand source and a stated obligation, from event metadata rather than a workshop. Where the evidence is thin the agent says so.
Demand3 readers / 41 recipients
Verified obligationnot located
Annual cost€186,000
Capacity draw2,240 h
03 · Challenge · weeks 4 to 6
Test why each activity exists
Agents test the reason behind each activity. Where owners disagree the contradiction stays visible rather than being averaged away, and anything the evidence cannot settle stops at a named reviewer.
F. Bildiriciconflict check clear
Recommendationredesign · 0.87
Contradictionboard requirement asserted, no rule located
04 · Decide · weeks 6 to 8
Record a decision the organisation keeps
Remove, redesign, automate or keep, each with an owner, a reversibility class, the evidence it rested on and a date it will be looked at again. The workspace stays live after week eight.
DecisionRedesign
Decided byF. Bildirici · 12 Sep
Experimentproduce on request, 90 days
Revisit17 Oct
If a function does it the same way every month, ZeroBase can review it.
These are starting scopes. Every one is bounded to how your function actually runs.
Month-end close
Establish which parts of the close are still required, what each costs a year, and which rule requires them. Never infer an obligation from precedent.
Start a Clean Sheet Review →- 1Sources connectedGL, close calendar, reporting portal
- 2Inventory established248 activities · cost, capacity, demand
- 3Obligations tested2 verified · 7 asserted, no rule
- 4Gate opened3 recommendations to named reviewers
September close · Finance34 recurring activities · 7 with no rule located
Activities inventoried: 248
Obligations verified: 2 of 9
Reviewers notified: S. Chen + F. Bildirici
StatusReady for gate
Approval chains
Trace every step in the chain, count what each one has actually rejected, and find the value at which it stops paying for itself.
Start a Clean Sheet Review →- 1Chain traced4 approvers · 1,120 requisitions
- 2Rejections counted0 below €5,000 · 18 months
- 3Rule locatedthreshold policy, 2019
- 4Change proposedtwo steps above €5,000, reversible
Requisition approval chain · CTL-0120 rejections below €5,000 in 18 months
Steps traced: 4
Rejections found: 0
Capacity: 1,940 h a year
StatusRedesign proposed
Ticket routing
Compare the documented route with the one tickets actually take, count the handoffs the gap produces, and make the real path the standard one.
Start a Clean Sheet Review →- 1Tickets read4,180 · metadata only
- 2Paths compared29% documented · 71% exception
- 3Handoffs counted3.2 per ticket · 88% re-keyed
- 4Route draftedexception path becomes standard
Access change request · REQ-11471% of 4,180 tickets on the exception path
Handoffs per ticket: 3.2
Fields re-keyed: 88%
Capacity: 1,510 h a year
StatusAwaiting reviewer
Inspection intervals
Test whether an interval is set by risk or by habit: what it has found, what it costs, and whether condition data already answers the question.
Start a Clean Sheet Review →- 1Work orders read4 cycles · 212 assets
- 2Findings counted0
- 3Condition data checkedtelemetry on 198 assets since 2024
- 4Experiment proposedcondition-triggered · revert on first finding
Quarterly asset inspection · OPS-0170 findings in 4 cycles · condition data on 198 of 212 assets
Findings, 4 cycles: 0
Condition telemetry: available
Capacity: 1,120 h a year
StatusExperiment proposed
Recurring reporting
Find who actually reads each recurring report, what producing it costs, and whether the rule that requires it can be located at all.
Start a Clean Sheet Review →- 1Access logs readreporting portal · 12 months
- 2Readers counted3 recurring of 41 recipients
- 3Rule searchedpolicy register · 41 documents · none
- 4Gate openedG-02 · experiment requested
Monthly performance pack · ACT-0413 readers of 41 recipients · €186,000 a year
Readers, 12 months: 3
Rule located: none
Owner claim: board requirement
StatusAt gate · G-02
Supplier reviews
Compare the last six review cycles, count what changed, and decide whether a quarterly cadence is still a control or only still a calendar entry.
Start a Clean Sheet Review →- 1Cycles compared6 · 41 suppliers
- 2Findings diffed0 changed
- 3Capacity attributed1,050 h a year
- 4Change proposedannual cadence · exceptions on trigger
Quarterly supplier review · CTL-031findings unchanged in 6 cycles · 41 suppliers
Cycles compared: 6
Findings changed: 0
Capacity: 1,050 h a year
StatusRedesign proposed
04 · What you work in
Creative ideation
Would you build it again?
The clean sheet is the signature surface. Each activity goes into one of three columns and the cost and capacity delta moves with it. Not an optimisation of today’s process. The question is whether it would exist if you designed the function again.
04 · What you work in
Data visualization + analysis
Every figure, traced to its locator.
A recommendation opens into a typed graph: the claims behind it, the system, query and period each came from, and the contradiction that stays visible instead of being folded away. This is the ontology: activities, obligations, evidence, gates and decisions as first-class objects with named relations.
04 · What you work in
Real-time collaboration
One workspace, named reviewers.
Function owners, named expert reviewers and the agents work one activity inventory. A reviewer can approve, challenge an assumption, request evidence, request an experiment, redirect or escalate, and everyone sees the same state, with the trail attached.
04 · What you work in
Planning
Current to target, reversibly.
Every decision becomes a delta with an owner, a reversibility class, the experiment that tests it and the date it is looked at again. The plan is the set of deltas the evidence supports, not a roadmap written before the evidence existed.
04 · What you work in
Purpose-built productivity suite
Decisions the organisation keeps.
Producing a recommendation is becoming a commodity. Holding why an organisation decided something, on what evidence, under which conditions and when it will look again is not. Decision memory is where reviews compound instead of restarting.
05 · Your stack
Works with the stack you already have.
SAP, Oracle, Workday, ServiceNow, Jira, Microsoft 365, Slack, Salesforce and the internal systems around them. Nothing is replaced, and nothing is written to.
- Read-only connectors where a source exposes event metadata
- Metadata only: never document content, never a measurement below the activity level
- Closed environments: cloud, private cloud, on-premise or isolated, with no outbound dependency
- Microsoft 365
Slack
Salesforce
- CLOUD
- PRIVATE CLOUD
- ON-PREMISE
- ISOLATED
Compatible source categories and deployment targets. Connector coverage is in design-partner validation; no live production integration is claimed.
06 · Maintenance
We carry the maintenance.
ZeroBase sits between the systems you already trust. Sources change, thresholds move, reviewers rotate. The workspace keeps running and the memory keeps compounding.
Start a Clean Sheet Review →No rip-and-replace
Keep the ERP, the ITSM, the document tools and the approval process your team already uses. ZeroBase reads them; it does not replace them.
No key-person risk
Decisions live in a workspace with their evidence and history, not in one analyst’s spreadsheet or a consultant’s deck.
No maintenance burden
When a source changes or a threshold needs to move, the review keeps running and the change is recorded with the person who made it.
Gated by design
No consequential change goes live without a named reviewer, and anything not reversible waits at the gate before it moves.
A generic agent cannot pass your audit. ZeroBase can.
Models read, classify and draft. They never decide what your operation does.
| A generic AI agent | ZeroBase |
|---|---|
| Decides what to do at runtime and can invent a figure or a rule. | Reads approved sources within a declared scope. Anything unlisted is denied. |
| Averages a contradiction into a confidence score. | Keeps a contradiction visible until a named person resolves it. |
| Hides its reasoning inside the model. | Keeps every claim with the system, query and period it came from. |
| Asks you to trust that the right thing happened. | Leaves a decision record with its evidence, its owner and its review date. |
Start with one function.
Eight weeks from baseline lock to recorded decisions. About 90 minutes a week from the function owner and sponsor.
A Future Innovation Agents product