Custom agents thatdo real work.

We design, build and run AI agents and models for your business. From the first data audit to the 3 a.m. alert. One team, one contract.

Book a call See the process

Agents in production at

Norvale Freight
Halden Mutual
Brightwater Clinics
Okapi Retail
Fennick Energy
Quarry & Lane
Ostra Bank
Meridel Foods

Most AI pilots stop at the demo.

A demo answers ten questions well. Production answers ten thousand, some of them badly. The gap is data, tests and operations. That gap is our job.

71%

of pilots never reach production*

4.6 months

average stall before a pilot is shelved*

eval suites in most stalled pilots*

*Tessera intake survey, 212 companies, 2025.

Seven parts. One system.

You can hire us for one part. Most clients take all seven.

01
Strategy
We find the 2–3 tasks where an agent pays back in a year.
Output: a ranked use-case list with cost and risk.
02
Data preparation
We clean, label and version your data.
Output: a dataset card and a reproducible pipeline.
03
Fine-tuning
We adapt open or hosted models to your domain.
Output: model weights you own, plus a training report.
04
Agent building
We connect the model to your tools: CRM, ERP, ticketing, email.
Output: an agent with typed tool calls and clear limits.
05
Evaluation
We write the tests before launch.
Output: an eval suite with pass rates per task.
06
Deployment
We ship to your cloud or ours.
Output: a versioned release with rollback in one command.
07
Monitoring
We watch cost, quality and drift every day.
Output: a weekly report and an on-call rotation.
Start with one partBook a call

Ten weeks to a working agent.

Each phase has a gate. We do not move on until the gate passes. You see every gate result.

Week 2Use case signed off. Cost model agreed.
Week 4Dataset v1 frozen. Label agreement ≥ 0.85.
Week 7Model beats baseline on eval suite.
Week 9Agent passes 95% of task tests in staging.
Week 10Live on 10% of traffic. Rollback tested.

Every step is logged.

A Tessera agent is not one prompt. It is a router, tools, a model and a guard. Each step writes a trace. You can replay any answer.

Trace retention90 days
Max tool calls per task12
Guard checksPII, policy, cost
Human hand-offon low confidence

Measured, then shipped.

agents in production

2.1M

tasks per month

94.2%

median task pass rate

11 weeks

median time to production

33-Point Accuracy Gain. 94.4% pass rate vs 61% prompt-only. Claims-handling agent, Halden Mutual, 2025. 412 eval tasks.

Pass rate is the share of eval tasks the agent completes with no human fix. Numbers from a claims-handling agent, Halden Mutual, 2025.

We pick the model by test, not by name.

45× Cost Reduction. $0.90 vs $41.00 per 1k tasks. Same 412-task suite. Picked model in gilt.

We run every candidate model on your eval suite. Then we pick on accuracy, cost and speed together. Sometimes the small fine‑tuned model wins.

Model shortlist, claims triage task (sample)
Model (anonymised)	Type	Task pass rate	Cost / 1k tasks	p95 latency	Runs in your cloud
Model A	Large hosted	93.1%	$41.00	4.8 s	No
Model B	Mid hosted	89.4%	$9.20	2.1 s	No
Model C	Open 70B, base	84.7%	$6.10	3.4 s	Yes
Model C-ft	Open 70B, fine-tuned by Tessera	94.6%	$6.10	3.4 s	Yes
Model D-ft	Open 8B, fine-tuned by Tessera	91.8%	$0.90	0.7 s	Yes

Picked: Model D-ft. 2.6 points below C-ft, 85% cheaper, 5× faster.

Three agents. Three results.

Logistics, Antwerp — 2025

Norvale Freight

Task: read carrier emails, update shipment status in the TMS.

Manual updates −82% in 4 months

Emails / day: 6,400
Pass rate: 96.1%
Hand-off to human: 3.9%

Read case

Healthcare, Leeds — 2024

Brightwater Clinics

Task: draft referral letters from visit notes. A doctor signs every letter.

Letter time 14 min → 3 min

Letters / month: 11,200
Edits per letter: 1.4
Data region: UK only

Read case

Finance, Tallinn — 2025

Ostra Bank

Task: first-line answers to card disputes, with policy checks.

First reply time −71% in 6 months

Disputes / month: 38,000
Policy guard blocks: 0.6%
CSAT: 4.4 / 5

Read case

Launch is week one.

64% Cost Reduction. 10.4 → 3.7 cents per task in 12 months. Pass rate up to 95.6%. Card-dispute agent, Ostra Bank.

Models drift. Prices change. Your process changes. We watch quality and cost every day, and we retrain when the tests say so. Not on a calendar.

M3Cache added. Repeat questions stop paying for a model call.
M6Retrain moved the agent to a smaller model at the same pass rate.
M9Drift alert, 620 new labels, retrain. Pass rate back above 95%.

Tessera vs. the usual options.

Tessera compared with building in-house and a generic AI platform
Criterion	Build in-house	Generic AI platform	Tessera
Time to first agent in production	6–12 months	2–4 months, limited tasks	10–12 weeks
Custom fine-tuning	If you hire for it	No	Yes, you own the weights
Eval suite before launch	Rare	Generic benchmarks	Written for your tasks
Runs in your cloud	Yes	No	Yes
Monitoring and on-call	Your team	Dashboard only	Our team, 24/7 on Run plan
Exit cost	Low	High (lock-in)	Low: code, data and weights handed over

Up to 5× Faster to Production. 10–12 weeks vs 6–12 months in-house. No lock-in.

Fixed scope. Fixed price.

Every engagement starts with a paid two-week Discovery. You keep the report if you stop there.

Discovery

2 weeks

€18,000fixed

Use-case ranking
Data audit
Cost model
Go / no-go report

Start Discovery

Build

from €140,000fixed · 10–12 weeks

One production agent
Data pipeline
Fine-tuning if needed
Eval suite
Deployment to your cloud
Code and weights handed over

Plan a buildmost chosen

Run

monthly

from €9,500per month

Daily monitoring
Monthly retrain if tests drop
On-call
Quarterly review
Cost tuning

Ask about Run

Run plan · service levels

Run plan service levels, Standard and Critical
Service level	Run Standard	Run Critical
Uptime (Tessera-hosted)	99.9%	99.95%
Sev-1 response	1 hour, business hours	15 min, 24/7
Sev-2 response	4 hours	1 hour
Quality alert threshold	pass rate −3 pts	pass rate −1.5 pts
Retrain turnaround	10 working days	5 working days
Reports	Monthly	Weekly
Price	from €9,500 / mo	from €21,000 / mo

15-Minute Sev-1 Response. 24/7 on Run Critical, with 99.95% uptime.

Questions, answered.

01Who owns the model and the code?

You do. At the end of Build we hand over code, data pipeline and model weights. No licence fee.

02Does our data leave our cloud?

Not unless you say so. We work in your AWS, Azure or GCP account. We can also work on-premises.

03Which models do you use?

Whichever passes your eval suite at the lowest cost. Hosted or open. We have no reseller deals.

04What if the agent gets an answer wrong?

The guard catches policy and data errors. Low-confidence tasks go to a human. Every answer has a trace, so we can find the cause.

05How much data do we need?

For most tasks, 500 to 5,000 good examples. Discovery tells you if you have enough.

06Can we start with one team?

Yes. Most clients start with one task in one team, then expand.

07Do you work under GDPR and the EU AI Act?

Yes. We provide a data processing agreement, a risk classification and model documentation for each agent.

08What happens if we stop the Run plan?

We do a two-week hand-over to your team. You keep the monitors, the eval suite and the runbooks.

Bring one task. We bring the rest.

Tell us the task you want to hand to an agent. We reply within one working day with a yes, a no, or three questions.

Custom agents thatdo real work.

Agents in production at

Most AI pilots stop at the demo.

Seven parts. One system.

Strategy

Data preparation

Fine-tuning

Agent building

Evaluation

Deployment

Monitoring

Ten weeks to a working agent.

Every step is logged.

Measured, then shipped.

We pick the model by test, not by name.

Three agents. Three results.

Norvale Freight

Brightwater Clinics

Ostra Bank

Launch is week one.

Tessera vs. the usual options.

Fixed scope. Fixed price.

Discovery

Build

Run

Questions, answered.

Bring one task. We bring the rest.