Scale the work. Keep the judgment.
Agents take the reading, matching and drafting. People keep every decision that affects a person, and sign for it.
A field guide from 8thElement
AI that moves the number, not the slide count.
Why most enterprise AI never reaches the P&L, how agents and people share the work when it does, and the playbook we run to get a named problem to a measured result in twelve weeks.
Scroll
Adding AI to a process that was never designed for it is not the same as changing the process. The firms that get value treat AI as a process decision, made by people who have run it.
What we believe
Three things we hold to on every engagement.
Make it traceable, or it does not ship.
Every action logged, every claim cited, every stage of autonomy earned on evidence a regulator can follow.
Measure the number, and tie the fee to it.
A baseline your finance team signs before we start, and a fee that moves only when that number does.
01
Why pilots stall
The model was rarely the problem. Neither was the data.
AI makes reading, matching and drafting cheap. It does not change who owns the result, who signs the decision, or who pays to keep it running. That is where pilots die.
The three failures we see most
Nobody owned the number
The pilot proved a capability. It never had a sponsor with a baseline to move, so there was nothing to be measured against and no one to be disappointed.
Governance arrived afterwards
Risk and compliance met the model at the end, when the only options left were delay or stop. Governance has to be in the design, not in the review.
Nothing was funded to run
The pilot budget ended. Nobody had the money, or the people, to keep it alive in production. A result that cannot be run is a slide.
43 initiatives. Ten carried on.
Nineteen stopped, fourteen merged into six, ten accelerated. $7.5M+ in annual savings came from stopping work as much as from starting it.
Read the case →Let us pause. Which of the three do you recognise?
02
Agents and people
Agents do the reading. Your people judge and sign.
A prior authorisation nurse was hired for judgment, and spends the day assembling evidence. Give the assembly to an agent and the judgment gets the day back.
The operating model
Three layers. One accountable team.
Layer 01 · Agents
Agents handle the volume
Reading, matching, assembling, drafting. Always consistent, always available, always ready to be tuned.
Layer 02 · Your experts and our operators
People own the judgment
Edge cases, overrides, every decision that affects a person. The signature a regulator will ask about.
Layer 03 · The sponsor and finance
Leadership owns the number
What was accepted, what was corrected, what moved. The loop that turns a result into the next one.
So instead of asking
- Which model is best?
- Can we prompt it better?
- How smart is it?
Ask instead
- Who owns the number it is meant to move?
- Can a regulator trace the decision?
- Who signs, and what happens when they say no?
How an agent earns the right to act
Autonomy is earned in stages, never assumed.
- Stage 1
Shadow
Output compared. Your team decides exactly as before.
- Stage 2
Approve
The agent recommends. A named person decides on every output.
- Stage 3
Bounded autonomy
Acts within limits your risk owner sets. Everything outside escalates to a person.
Adverse decisions about a person are never made by an agent. Every action is logged and reversible, and your team can pause any agent in one action.
How is this sitting with you so far?
03
How an engagement runs
AIM: assess, implement, mature. Start at whichever fits.
You do not switch AI on. You graduate it, one named process at a time, with a gate you can check at the end of each phase.
Assess
Where do we actually stand? A ranked backlog: every use case with a sponsor, a baseline and a target.
Implement
Can we ship one outcome and prove the pattern? An agent in approve mode with a measured business result.
Mature
Can it run better with fewer hands on it? Your team runs it, and we move from running to advising.
Most clients start where the gap is and complete the full cycle in six to nine months.
Proof
Five engagements, five numbers someone owned.
$0annual savings, AI portfolio consolidated
0clinical guidelines per reviewer day
$0indirect spend under a learning model
$0found against a $40M target
0from signed spec to measured outcome
Results are described without naming the client.
04
The playbook
The part that gets left out of the brochure.
Six plays, in order
Start boring
Pick a workflow that is high volume, bounded, repeatable and reviewable. Your most ambitious use case is not the first one. Boring is where you can define and measure success.
Sign the baseline
Before anything is built, the sponsor and finance agree what the number is today and what it should be. Without that signature there is nothing to move.
Build in shadow
The agent runs against live work while your team still decides everything. You learn what it gets wrong before anyone depends on it.
Check every output early
Early on, people review all of it. That review is training as much as quality control. Oversight reduces only when the evidence says it has been earned.
Measure the number, not the tokens
Throughput, acceptance rate, cost per outcome. Argos meters every model call so finance can read the cost line next to the result.
Expand and compound
What worked becomes a playbook. The next workflow starts ahead, autonomy expands on evidence, and the programme stops being a project.
How AIM runs →One more. What would need to be true for you to start?
Take it with you
Get the field guide as a PDF.
Eight pages, laid out for printing and sharing: the three failures, the operating model, how an engagement runs, the proof, and the playbook. Leave a work email and the download unlocks here.
- Work email only; we reply to company addresses
- We use your details only to send the guide and to reply to you
- An operator replies within one business day if you ask us to
Where you put AI, who signs for it and how it is measured decide whether it becomes a new source of risk or the thing your operation has needed for years.
Where this leads
Tell us the number you need to move.
A forty five minute working session with an operator who has run the kind of work you are describing, and an honest read on what it would take.