Agentic Decision AI
Contact

AD Tech AI Lab · Toronto

Find the best AI model for your work: the accuracy you need at the lowest cost.

We turn your apps into a test environment, write real-world scenarios, and measure frontier and open-source models on accuracy and cost.

Talk to usSee how it works

Inside our first environment

Synthetic patients with years of history, spread across seven connected apps. Encounters, orders, lab results, referrals and claims reconcile with one another. Documents arrive as PDFs, spreadsheets, faxes and scans, so a model has to read as well as act.

of synthetic patients
Thousandsof synthetic patients
documents across 15 types
4,600+documents across 15 types
scanned and faxed pages
1,000+scanned and faxed pages
Are you interested in buying?

Accuracy vs cost, per task

Schematic, not data
Schematic of accuracy against cost per taskA scatter of frontier and open-source models with a curve along the best trade-offs. A horizontal line marks an accuracy bar. The best option is marked where the curve first clears the bar: accurate enough, at the lowest cost. The positions are illustrative and say nothing about any real model.Cost per task →Accuracy →Your accuracy barBest foryour bar
  • Frontier models
  • Open-source models
  • Best trade-offs
  1. 1

    Your apps become a test environment

  2. 2

    Real-world scenarios are written

  3. 3

    Frontier and open-source models run them

  4. 4

    You get the best option for your bar

A schematic of what we deliver. Positions are illustrative, and no real model or result is shown.

The problem

Teams choose AI models from general leaderboards and price lists.

Neither says how a model handles your apps and your workflows. The top model on a leaderboard may cost too much at your volume, and a cheap one may fail the cases that matter. You only find out by testing on your own work.

How it works

Four steps from your software to a decision.

  1. 01

    Your apps become an RL environment

    We build a working test version of your software, wired to the data that represents your work. Models act in it the way your staff would.

Schematic · illustrative
Clinic softwareRecords viewerDocument intakeReferralsClaimsPatient portalShared hubPDFFAXXLSXSCANCSVWHAT CHANGED IN THE RECORDS+++Cost per task →Accuracy →Your accuracy barBest for your barBEST

What you get

Accuracy and cost, side by side, for every model.

Each evaluation ends in a report like this one. The values below are samples that show the layout. They are not measured results for any model.

Drag your accuracy bar

The best fit, the table and the savings update together

Sample values · not results
60%70%80%90%100%$0.02$0.10$0.50$1Cost per task →Accuracy →GPT-6 Astra: 94% accuracy, $1.10 per task, $1.17 per correct resultGemini 4 Carbon: 93% accuracy, $0.95 per task, $1.02 per correct resultClaude Fable 5.2: 91% accuracy, $0.84 per task, $0.92 per correct resultGemini 4 Argon: 90% accuracy, $0.62 per task, $0.69 per correct resultMuse Wonder Full: 89% accuracy, $0.70 per task, $0.79 per correct resultClaude Opus 5.5: 87% accuracy, $0.52 per task, $0.60 per correct resultGrok 4.8: 86% accuracy, $0.38 per task, $0.44 per correct resultMuse Wonder: 85% accuracy, $0.33 per task, $0.39 per correct resultMuse Spark 1.3: 80% accuracy, $0.22 per task, $0.27 per correct resultMuse Spark 1.3Gemma 4: 69% accuracy, $0.03 per task, $0.04 per correct resultQwen3.5-27B: 64% accuracy, $0.02 per task, $0.03 per correct resultYour accuracy bar · 75%

Best fit at a 75% bar: Muse Spark 1.3, $0.27 per correct result.

Model evaluation report

Your environment · your scenarios

Sample values · not results

Best for your 75% bar

Muse Spark 1.3

80% accuracy · $0.22 per task · $0.27 per correct result

How “best” is decided

The lowest cost per correct result among the models that clear your accuracy bar. Change the bar and the answer can change.

Sample values for layout only. These are not measured results for any model.
RankModelTypeAccuracyCost per taskCost per correct resultForbidden actionsHarness errors
1Muse Spark 1.3BestFrontier
80%
$0.22$0.2700
2Muse WonderFrontier
85%
$0.33$0.3900
3Grok 4.8Frontier
86%
$0.38$0.4400
4Claude Opus 5.5Frontier
87%
$0.52$0.6000
5Gemini 4 ArgonFrontier
90%
$0.62$0.6900
6Muse Wonder FullFrontier
89%
$0.70$0.7900
7Claude Fable 5.2Frontier
91%
$0.84$0.9200
8Gemini 4 CarbonFrontier
93%
$0.95$1.0200
9GPT-6 AstraFrontier
94%
$1.10$1.1700
–Gemma 4Below barOpen-source
69%
$0.03$0.0400
–Qwen3.5-27BBelow barOpen-source
64%
$0.02$0.0300

Sample values to show the layout. They are not measured results for any model. Real reports carry a 95% confidence interval, and forbidden actions and harness errors are reported separately from ordinary failures.

Why the choice matters

Don't follow a model. Choose the one that fits your work.

The gap between the default pick and the best fit can be large. Here is what it looks like at your volume.

  • A leaderboard is not your workload

    General rankings test general tasks. Your apps, data and edge cases are on no leaderboard, so the top model there can be the wrong one for you.

  • Prices differ several-fold

    Frontier models can differ several times over in price per token, and accuracy does not rise in step. Paying for the top model on work a cheaper one handles is money lost.

  • Cheap can fail where it counts

    A low price means little if the model fails the cases that matter. Every wrong record is rework for a person.

  • The field moves every few weeks

    New frontier models now arrive every few weeks across the major labs. The best fit today may not be next quarter's, so measure again instead of assuming.

What the right choice can save

Illustration · edit the inputs

The starting values come from the sample report above: the most accurate model against the best fit for a 75% bar. They are not measured results.

Choosing the right model can save you approx. US$540,000 a year.

A month, approx.
US$45,000
Lower cost per correct result
77%

At 50,000 tasks a month, $1.17 against $0.27 per correct result.

An illustration, not a promise. Your real figure comes from an evaluation on your own apps, workflows and volume.

We've built one

A seven-app clinic, built end to end.

Our first environment is a simulated clinic: seven connected apps that share one patient record, a safety gateway that blocks forbidden actions and logs every one, and a grader that reads the records after each run. Healthcare comes first.

  • Graded on final state

    A pass needs the records to end up correct and no forbidden action in the log.

  • Every case has a twin

    One decisive fact changes, or the evidence is removed, so a lucky guess does not pass.

  • Escalating is valid

    When the evidence is missing, handing the case to a person is graded as correct.

  • Read the research

Our first product

Live now

Iris

An AI customer support employee for e-commerce. Iris answers order questions, verifies identity and handles refunds and invoices, with the owner in control.

Book an Iris demoMeet Iris

Get in touch

Choosing an AI model for your work?

Tell us about your apps and workflows. We will show you how we would measure which model is the best fit, balancing accuracy and cost.

Talk to usFollow our research