Skip to main content
Polymath DigitalStart with one workflow

ServicesAI pilot to production

Our AI pilots work. Why aren’t they reaching production?

Because a pilot runs on the documented path and production runs on the actual one. An agent can only act where someone has written down who owns the decision, what the role may approve and access, and when a person must step in. Those answers exist in your organization, but rarely in one place.

What it looks like

  • The pilot handles the standard case and stops at the first exception
  • Nobody can say which system governs when two of them disagree
  • The agent can reach more data than the person it acts for
  • Approval authority lives in a policy document rather than in a system
  • Each function has run its own pilot, on its own definitions
  • The business case cannot name the process the agent would finish

Why it happens

  • The actual path was never written down

    Processes have a documented path and an actual path, and only one is written down. The judgment behind exceptions, price mismatches and one-off approvals sits with experienced people, which is exactly what a model cannot supply.

    AI in the Enterprise

  • Role and authority are defined in several places

    An agent inherits the permissions of the role it acts for. When positions, approval limits and access are maintained separately by HR, finance and identity, the agent inherits whichever version it happens to reach.

    Hire to Retire Runs on One Event

  • The workaround is the real specification

    The spreadsheet and the chat thread running alongside the process are not defiance. They are the specification nobody wrote down, and automating the documented process leaves them in place.

    Bottom-Up Transformation

How we approach it

We start with one workflow: a process that crosses three or more functions, mapped end to end with the people who own its steps.

  1. Choose the process by where the exceptions cost you, not by where the volume is.
  2. Write the actual path by interview, starting with whoever handles what goes wrong.
  3. For every decision an agent would take, answer four questions: who owns it, what authority applies, which data is authoritative, and when a person steps in.
  4. Name one author for each field, and give each fact back to its source.
  5. Route approvals to the constraint rather than to the hierarchy.

What you get

  • One process written as it is actually performed, exception path included
  • Four written answers for every decision an agent would take
  • A role model the agent can inherit, rather than a permission set maintained beside it
  • The share of transactions completing the documented path with no deviation — expect a number well below the one you would guess
  • How many steps produce two different answers when two experienced people are asked which source governs

Related perspectives

  • Paper 14

    AI in the Enterprise

    Your processes have a documented path and an actual path. Only one is written down

    Read

  • Paper 15

    Hire to Retire Runs on One Event

    The hire happened once. Whether it is entered once is an architecture decision, not an HR one.

    Read

  • Paper 12

    Bottom-Up Transformation

    The workaround is not defiance. It is the specification nobody wrote down

    Read

  • Paper 17

    Letting the Person Who Knows the Fact Record It

    Automation here does not remove work. It moves authorship to the only person who cannot get it wrong

    Read

  • Paper 16

    One Demand Signal, Four Currencies

    Your four forecasts are not four opinions about one number. They are four different quantities

    Read

  • Signal

    Workday Rising 2026: Sana Works Across Your Stack. Your Organization Model Decides How Far.

    Rising puts Sana for HR, Finance and IT on the main stage. How organizations running Workday alongside SAP, Oracle, Salesforce or ServiceNow can get the most from it.

    Read

Questions we are asked

Why don’t enterprise AI pilots scale?
A pilot is scoped to the path the documentation describes. Production meets the exception path, where the judgment lives with experienced people and has never been written down. The gap is not model capability; it is that nobody has recorded who decides what.
What makes a process AI-ready?
Every decision in it has four answers written down: who owns the decision, what authority applies, which source is authoritative when two disagree, and when a person must step in. A process with those answers can be acted on; one without them can only be described.
How do roles, approvals and access affect AI agents?
An agent acts with the authority of the role it acts for, so the role model becomes the control on what it can do. Approval limits and access defined once, and inherited by every system, give every agent the same answer to who may do what.
Which AI platforms do you work with?
We are platform-neutral. Most organizations will run agents from more than one vendor — Google Gemini, Microsoft Copilot, Claude, Sana from Workday — and none of them can decide on its own who in your organization holds authority. That answer comes from your organization model, whichever platform reads it.
Where do you start?
With one process where exceptions are expensive, and the people who handle those exceptions. We write the actual path, then the four answers each decision needs.

Value. Realized.

Start with one workflow.

Choose one process that crosses three or more functions. We map it end to end with you, and mark every place the same problem gets solved twice.