How we work

Five stages, each with a duration, a price band and a way to stop

Most agencies publish a four-arrow diagram — discover, design, deliver, support. It is true of every firm on earth, which is what makes it useless to you. Below is the sequence we actually sell: the fee band and duration of each stage, the artifact you keep at the end of it, the decision it forces, and the condition under which we stop — including the conditions that end the engagement against our own interest.

Book a call See our projects →

Two people at a desk reading an extracted process log on screen during a Palamed baseline stage.

30 min: first call, no charge, with the engineer who would build itYou get a written note within two working days naming the three processes we would measure first and the one we would leave alone. No obligation, and no proposal attached unless you ask for one.
2–3 weeks: from signature to a baseline that is yours to keepFixed fee, EUR 6,000–14,000 depending on how many systems the event log has to be reconstructed from. Credited in full against a build that starts within three months.
80%: data-readiness floor below which we stop rather than buildShare of cases carrying complete timestamps and resolvable identifiers. Below it, mining results are unsafe and we hand over a remediation list instead of a proposal. Our own gate, not an industry standard.
4 business hours: MTTR we put in the contract for a broken automationA back-office target we design and report monthly against — not a measured historical average, because we will not publish an average we cannot show you the log for.

We start with a stopwatch, not a demo

Instead of asking what you would like automated, we ask what the process costs you today, measured the same way twice. That one change decides everything downstream: it produces a baseline your finance team will accept as a denominator, it re-ranks the backlog, and it usually retires one or two of the candidates you arrived with.

The baseline is reconstructed from system truth — change documents and audit tables, not interview notes. In SAP that is CDHDR and CDPOS joined to EKKO/EKPO/EKBE or BKPF/BSEG; in Dynamics 365 Business Central, Odoo, or a Bulgarian package such as Microinvest or Ажур L, the equivalent audit tables. We report p50 and p90 throughput separately, because the average is the number that hides the problem you actually have.

The person who runs that extraction is the person who was on your first call and the person who writes the integration. We are a founder-led studio with a network of named specialists, so there is no account layer between you and the work — and no junior learning on your month-end close.

What we need from you

A stage cannot be honest about its duration if the inputs arrive late. These are the four things that actually move our dates, and we would rather name them before you sign than explain them in week three.

Read access to the source tables, or an extract of them, with a named person in IT who can grant it. This is the single most common reason a two-week baseline becomes a five-week one.
One process owner with the authority to decide an exception rule, available roughly two hours a week. Not a steering committee — a person.
A non-production environment to write into, or an honest statement that none exists. If none exists we will say what that costs in caution and in calendar time, rather than pretending it does not matter.
The benefit definition agreed with finance before the build, in writing: which of headcount not backfilled, contractor spend removed, early-payment discounts captured, penalties avoided or DSO days moved actually counts. Six minutes saved across forty people never appears in a P&L, and we will not claim it.

What you keep whether or not we continue

Every stage terminates in something that is yours, in a format you can hand to another supplier. That is deliberate. An engagement that only produces value if it continues is a hostage arrangement, and it is why so many discovery reports read as sales documents with a page count.

The written note from the first call — the three processes we would measure first and the one we would leave alone — sent within two working days, whether or not you ever reply.
The Process Ledger: the extracted event log itself in XES or OCEL 2.0, the variant analysis, p50 and p90 per variant, rework rate per activity, and the ranked backlog. Portable into Celonis, Signavio, Apromore or PM4Py — not locked to us.
The pilot code, its tests and its configuration, in your repository from the first commit rather than transferred at the end. We have never held a client repository and we do not want the option.
Runbooks, the data-flow map naming every place personal data lands, the decommission plan, and the change-control process itself. The last one matters most: if your team cannot ship a change to the automation, we built a dependency.

How we price

Each stage is quoted as a fixed fee for a defined scope and invoiced against that stage alone. We do not price a whole programme up front, because we would have to price our own uncertainty into it and you would be paying for our ignorance.

The bands below are indicative — what a Bulgarian founder-led studio would expect to quote on a mid-market engagement, not a quote. They exclude VAT and third-party licence cost, which we model separately at three times your current volume so the licence bill is not a surprise in year two.

Baseline: EUR 6,000–14,000 for two to three weeks, depending on how many systems the event log has to be reconstructed from and whether a usable extract already exists.
Pilot on production data: EUR 14,000–38,000 over four to eight weeks. The baseline fee is credited in full against this stage if you continue within three months.
Build and handover: EUR 30,000–110,000 over eight to twenty weeks, driven by the number of integration surfaces and by whether a non-production environment exists.
Run: EUR 1,200–4,500 per month, rolling, thirty days’ notice either way, no minimum term. A retainer you cannot leave is a subscription to our convenience.
What moves a quote inside its band is almost never the model or the framework. It is how many systems have to agree, whether the target system has a released write interface, and how much of your document mix is degraded Cyrillic scan rather than digital PDF.

What we do not do

Four refusals, each of which has cost us work.

We do not quote a fixed price for an undefined scope. Naming a number before we have seen your event log means pricing our own uncertainty into your invoice, and the supplier wins that trade every time. We lose work to firms who will give you a confident number in the first meeting, and we would rather keep losing it.
We do not build a pilot we do not believe can reach production. A demo on synthetic data is easy, and it is how most AI projects buy themselves one more quarter of budget before dying quietly. If the constraint that will kill it in production is already visible at pilot stage, the honest move is to say so while it costs you EUR 20,000 rather than EUR 200,000.
We do not write to your ERP through unreleased tables, direct database updates or screen scraping. It works on the day, and then an S/4HANA upgrade, a Business Central release wave or a vendor UI change breaks it silently — and silent breakage in a finance process is worse than no automation at all. Where genuinely no interface exists, we will say UI automation is the only route, quote its break rate honestly, and leave the decision with you.
We do not pretend AI is always the answer. A cleaner supplier master, a corrected tolerance rule or a deleted approval step is frequently worth more than any model, and a substantial share of what we recommend contains no machine learning at all. That removes us from the shortlist of buyers who came specifically for AI, which is a filter we are content with.

Where the method comes from

None of this is ours. The fixed-appetite stage with a stop condition is Basecamp’s Shape Up circuit breaker: if a project runs over, by default it does not get an extension, because the alternative is spending multiples of the original appetite on a concept that needed rethinking first.

The event-log discipline comes from process mining as formalised by Wil van der Aalst at TU Eindhoven and standardised as IEEE 1849 (XES), with OCEL 2.0 for the cases where a single case ID cannot honestly represent one order that becomes three deliveries and two invoices. The durable-execution constraints — determinism, idempotency keys, compensating transactions — are Temporal’s and Camunda’s, not a framework we invented. The write-path rule is SAP’s own clean core doctrine.

We cite the lineage because a method with no ancestors is usually a sales artefact. These are the best ways we have found so far; the stage names have changed twice since 2023 and will change again.

How the work runs

01
First call
30 minutes, no charge
What you keep
A written note within two working days: the three processes we would measure first, the one we would leave alone, and the reasoning for both. Yours to circulate internally, with no proposal attached unless you ask for one.
The decision it forces
Whether there is a measurable process here at all, and whether one named person can decide an exception rule. If both are yes, we quote a baseline; if not, we say what would have to change before there is anything worth quoting.
When we stop
If the cost sits in an unowned decision, a reorganisation or a supplier relationship rather than in a system, we say so on the call and do not send a quote. Automating around an organisational problem makes it faster and less visible.
02
Baseline
2–3 weeks
What you keep
The Process Ledger: the event log extracted from your system tables in XES or OCEL 2.0, variant analysis, p50 and p90 throughput per variant, rework rate per activity, and a ranked backlog carrying the automation ceiling the variant tail implies. Typical investment EUR 6,000–14,000, credited in full against a build that starts within three months.
The decision it forces
Which two or three processes get built and in what order — and which of the candidates you arrived with are retired, with the volume and variant evidence for why.
When we stop
If fewer than roughly 80% of cases carry complete timestamps and resolvable identifiers, the log cannot support a defensible baseline. We hand over the ledger and a data-remediation list and stop there rather than building on a fiction.
03
Pilot on production data
4–8 weeks
What you keep
One process running end to end against production data in a shadow lane, with its tests, its typed exception queue, and a re-measurement using the same method as the baseline so the two numbers are comparable. Typical investment EUR 14,000–38,000. Not a demo on synthetic data, and not a slide.
The decision it forces
Go or no-go on production, taken against a threshold both sides wrote down before the build — usually a touchless rate and a p50 cycle time, with the residual exception rate stated as a range rather than omitted.
When we stop
If the re-measured number misses the written threshold, the build stage does not start. We report the mechanism that held it back instead of reframing the result, and you have spent one stage fee rather than a programme budget.
04
Build and handover
8–20 weeks
What you keep
The running system, in your repository and your infrastructure from the first commit: typed integrations against released interfaces, durable orchestration with idempotency keys and a dead-letter queue, a runbook per automation, OpenTelemetry traces into your existing monitoring, a data-flow map naming every place personal data lands, and a decommission plan. Typical investment EUR 30,000–110,000.
The decision it forces
Whether you own it. Hypercare ends when your own team has shipped a change to the automation while we watched — not on a date in the plan.
When we stop
If at the end of the agreed hypercare window your team still cannot make that change without us, we have built a dependency rather than an automation. We extend handover at our own cost until they can.
05
Run
Monthly, rolling, thirty days’ notice
What you keep
A one-page monthly report: break rate per hundred runs per automation, MTTR against the four business hours in the contract, oldest item in each exception queue, auto-acceptance rate at the current confidence threshold, and the benefit trend against the original baseline. Typical investment EUR 1,200–4,500 per month.
The decision it forces
Renew, resize or leave — decided monthly against the report rather than annually against a relationship.
When we stop
If break rate trends upward and the cause is a system we cannot change, we recommend decommissioning the automation — including when it is our own build and the recommendation ends the retainer.

When we are the wrong choice

You need something demonstrable in three weeks. Our first stage buys a measured baseline, not a working bot — if you already know exactly what to automate and only need hands, a delivery shop will be faster and cheaper than us, and we will say so on the call.
The business case is headcount removal. At a Bulgarian hourly labour cost of EUR 12.0 against an EU average of EUR 34.9 (Eurostat, 2025), labour arbitrage is a weak case here — the returns are in cycle time, error rate, control, and absorbing volume growth without hiring.
Nobody owns the process. If the exception rule cannot be decided by one named person, no automation survives its first month, and we would be selling you a change-management project wearing a technical costume.
You want a single supplier for strategy, build, run, hosting and licensing. We are a small studio and we deliberately stay replaceable — if you need one throat to choke across an estate, hire a systems integrator.

Questions we get asked

What do we actually pay for if we stop after the baseline?

The fee for that stage and nothing else. You keep the extracted event log in XES or OCEL 2.0, the variant analysis, p50 and p90 throughput per variant, the rework rates, and the ranked backlog including the reason each rejected candidate was rejected. All of it is portable: another supplier can load the same log into Celonis, Signavio, Apromore or PM4Py and carry on without us. We design the baseline to be worth its fee as a standalone document, because a discovery report that is only useful if you buy the next phase is a sales deck with an invoice attached.

Who owns the code, the models and the data when this ends?

You do, from the first commit. Work goes into your repository, or a repository transferred to you at the first milestone, under an IP assignment in the contract rather than a promise on a website. Third-party components stay under their own licences and we list them with versions, so your legal team reviews a bill of materials rather than a black box. Your data stays in your systems or a tenancy you control; we do not train anything on it and we keep no copy after handover beyond what the support contract requires. Where a hosted model provider is involved, it is named in the data-flow map along with the region it runs in.

What happens if the pilot number does not move?

Then the pilot did its job and the build stage does not start. Before any code is written we agree a threshold — usually a touchless rate and a p50 cycle time — and the method by which it will be re-measured, which is the same method used for the baseline so the two figures are comparable. If it lands short, we report the mechanism that held it back rather than reframing the result, and you have spent one stage fee instead of a programme budget. That is the entire point of gating it: the alternative is a production budget approved on the strength of a demo.

How do we get out mid-engagement if this is not working?

Each stage is a separate fixed-fee agreement, so the natural exit is simply not starting the next one: no termination fee and no notice period between stages. Inside a stage, either side can stop with fourteen days’ notice, you pay for work completed to that point, and everything produced so far is handed over in the state it is in. The run retainer is rolling with thirty days’ notice either way and no minimum term. We keep the exit cheap on purpose — a client who cannot leave stops telling you the truth about how the automation is performing.

Can we skip the baseline and go straight to a build?

Sometimes, and we will tell you which case you are in. If you already have an event log, or a process mining deployment with the case-ID question resolved, we read your data instead of extracting it and the stage shrinks to about a week of validation. If what you have is a workshop diagram and a shortlist, skipping means building on your assumptions and then presenting the result as our number, which is not something we can stand behind. The middle option we do offer is a one-week baseline scoped to a single process rather than the full ledger, at the bottom of the band.

You are small — what happens to us if you disappear?

It is the right question to ask a founder-led studio, and it is why handover is a stage rather than a closing slide. Everything runs in your infrastructure under your accounts, the code and its tests are in your repository, every automation has a runbook, and hypercare ends only once your own team has shipped a change while we watched. The specialists we bring in are named to you rather than presented as our staff, so you know who built what and can contract them directly if we are gone. Nothing we build needs a Palamed licence, a Palamed cloud account or a Palamed person to keep running.

Thirty minutes, and a written note either way

Bring one process you suspect is expensive and the systems it touches. You leave with the three things we would measure first, the one we would leave alone, and the reasoning — written down and sent within two working days, whether or not you take it further.

Book a call Make an inquiry →

The person on the call is the person who would write the integration. No deck, no discovery questionnaire, and no proposal unless you ask for one.