Lastenheft and Pflichtenheft for AI Projects: What Changes with AI

Germany's classic requirements documents assume you can specify system behavior in advance. Probabilistic systems break that assumption. Here is how to adapt both documents so they still do their job.

Elena Voss·Head of AI Delivery, Aiporate··8 min read·Share on XLinkedIn

Key takeaways

  • The Lastenheft states WHAT the client needs and why (owned by the client); the Pflichtenheft states HOW the contractor will deliver it (owned by the contractor) and becomes the acceptance baseline.
  • Classic "input X must yield output Y, always" clauses are unwritable for AI systems: model behavior is statistical, so acceptance criteria must be statistical too, quality thresholds on defined test sets, not guarantees per case.
  • A usable AI Pflichtenheft specifies the evaluation, not just the system: the test dataset, the metrics, the thresholds, and the procedure become contractual objects in their own right.
  • Fix behavior at the boundaries: hard deterministic rules for what the system must never do (safety, compliance) can and should stay classic; the statistical part covers quality, not permissions.
  • Replace one-shot acceptance with staged acceptance: spec checkpoints per phase (data, offline quality, pilot behavior, production stability) keep the documents honest as the system evolves.

The Lastenheft and Pflichtenheft are the backbone of German project contracting: the client writes down what they need, the contractor writes down how they will deliver it, and acceptance happens against the second document. The model works because classic software is deterministic, given input X, the spec can demand output Y, always. AI systems refuse that deal. A model that is right 94% of the time is a success or a lawsuit depending entirely on how the spec was written. This article explains both documents briefly, shows exactly where probabilistic behavior breaks the classic logic, and gives spec patterns that keep the documents' function, clear accountability, without pretending the system is deterministic.

The two documents, briefly explained

For readers outside German-style contracting: the pair splits requirements into a client-owned and a contractor-owned document, and that split is what makes it powerful, when it works.

Lastenheft (requirements specification)Pflichtenheft (functional specification)
OwnerClientContractor (approved by the client)
Question answeredWhat is needed, and whyHow it will be delivered, concretely
Level of detailGoals, requirements, constraintsSolution design, features, technical detail
Legal roleBasis of the tender / contract scopeBaseline for acceptance (Abnahme) and warranty claims
Classic assumptionRequirements are fully knowable up frontBehavior can be specified exactly and verified case by case
Lastenheft vs. Pflichtenheft at a glance

Where probabilistic systems break the classic spec logic

Every one of these breakpoints has produced real project disputes across the industry. They are worth reading as a checklist against your current draft.

  1. 1No per-case guarantees: "for input X the system returns Y" is unwritable when the same input can legitimately yield different, or occasionally wrong, outputs. Only aggregate statements over defined test sets are honest.
  2. 2Acceptance becomes measurement: Abnahme against an AI spec is an evaluation run, which means the test data, metrics and procedure must themselves be specified, versioned and agreed, or acceptance is arbitrary.
  3. 3Requirements shift with the data: what is achievable depends on data quality discovered during the project, so a spec frozen before data exploration is fiction with signatures on it.
  4. 4Behavior drifts after acceptance: models degrade as the world changes, so a spec that ends at acceptance ignores the phase where most AI risk lives; monitoring and retraining duties belong in the Pflichtenheft.
  5. 5"Error" needs a definition: a wrong classification within the agreed error rate is not a defect (Mangel) in the warranty sense, and the documents must say so explicitly, or every miss becomes a claim.

Writing acceptance criteria for model behavior

The pattern: deterministic clauses for boundaries, statistical clauses for quality, and a specified evaluation procedure that makes the statistical clauses testable. The table shows classic clauses next to their AI-ready rewrites.

Classic clause (breaks with AI)AI-ready clause
"The system correctly classifies incoming documents""Macro-F1 >= 0.90 on the frozen test set TS-v1.2 (5,000 documents, stratified, jointly approved); evaluation script and seed versioned in the repo"
"The chatbot answers customer questions correctly""On the agreed 300-question eval set, >= 92% of answers rated correct by the defined rubric and two independent raters; zero answers violating the blocked-topics list (hard criterion)"
"The system responds within 2 seconds"Keep it classic: latency, availability and cost-per-request stay deterministic clauses, p95 latency <= 2s under defined load"
"Defects will be remedied within warranty""A quality drop below threshold on the agreed monitoring metric for 7 consecutive days triggers the remediation process in section X; misclassifications within the agreed error rate are not defects"
Classic spec clauses rewritten for AI systems

Iterative spec patterns that keep the documents useful

The goal is not to abandon the Lastenheft/Pflichtenheft discipline but to phase it. These patterns are contract-compatible and used across the industry in various forms.

  1. 1Two-stage Pflichtenheft: a binding version after the data-exploration phase, not before; the pre-exploration contract covers only the exploration itself, with defined deliverables (data report, feasibility assessment, target metrics proposal).
  2. 2Corridor targets instead of point targets: the Lastenheft states a minimum acceptable and a target quality level; landing between them triggers a pre-agreed price/scope adjustment instead of a dispute.
  3. 3Frozen eval sets as contract annexes: test data, metric definitions and evaluation scripts are versioned annexes; changing them requires a change request like any other contract change.
  4. 4Staged acceptance gates: separate partial acceptance for data pipeline, offline model quality, pilot-phase behavior, and production stability after N weeks, each with its own criteria and its own sign-off.
  5. 5Living-document clause: a defined cadence (for example quarterly) where both sides review monitoring results against the spec and formally update thresholds, with a named owner on each side.

A filled-in miniature example

An illustrative excerpt from an AI-adapted requirements pair for a fictional invoice-processing project, showing how classic and statistical clauses sit side by side.

DocumentClause (condensed)Type
LastenheftManual invoice data entry (currently about 3 FTE, fictional figure) shall be reduced by at least 60% without increasing posting error ratesBusiness goal
LastenheftInvoices that the system cannot process confidently must route to a human queue, never to auto-postingHard boundary
PflichtenheftField-level extraction accuracy >= 97% on frozen test set TS-A (2,000 invoices, jointly sampled); confidence threshold tuned so that auto-posted invoices reach >= 99.5% field accuracyStatistical
PflichtenheftDocuments with confidence below threshold are flagged and queued within 5 seconds; no document is auto-posted without passing validation rules V1-V14Deterministic
PflichtenheftWeekly monitoring report on extraction accuracy against a rolling labeled sample; two consecutive weeks below 95% triggers the remediation processPost-acceptance
Illustrative excerpt: AI-adapted spec clauses, invoice extraction (fictional example)

Common mistakes in AI specs

  1. 1Signing a fixed-price, fixed-spec contract before anyone has looked at the data: the single most common root cause of AI project disputes.
  2. 2Writing quality targets without specifying the test set and procedure, which makes acceptance a negotiation instead of a measurement.
  3. 3Putting safety boundaries into the statistical part: what the system must never do belongs in deterministic clauses with zero-tolerance testing.
  4. 4Ending the Pflichtenheft at acceptance: no monitoring metrics, no drift thresholds, no named operator, the riskiest phase is left unspecified.
  5. 5Letting the contractor own the test data unilaterally: eval sets must be jointly approved and frozen, or the spec can be gamed from either side.

Frequently asked questions

Do we still need a Lastenheft and Pflichtenheft for an AI project at all?

Yes, arguably more than for classic software, because ambiguity about what counts as success is the main source of AI project disputes. What changes is the content: statistical acceptance criteria, specified evaluation procedures, and staged gates instead of one-shot acceptance.

Can a fixed-price contract work for an AI project?

For the exploration phase and for well-understood, bounded deliverables, yes. For end-to-end model quality on unexplored data, a fixed price against a frozen spec transfers a risk nobody can price, which is why two-stage contracts (exploration first, binding spec after) have become the practical standard.

Who should define the test dataset for acceptance?

Both sides jointly, and it should be frozen and versioned as a contract annex. A test set defined unilaterally by the contractor invites optimizing to the test; one defined unilaterally by the client invites unrealistic samples. Joint sampling with documented criteria protects both.

How do EU AI Act requirements interact with these documents?

For higher-risk use cases, documentation, logging, human-oversight and monitoring duties under the EU AI Act map naturally onto the Pflichtenheft: the spec is the right place to assign who implements and who operates each duty. Treating them as an afterthought after acceptance is the expensive way to comply.

Head of AI Delivery, Aiporate

Elena has spent 12 years building and embedding AI and data teams inside B2B SaaS companies, from first pilot to enterprise-wide platform. At Aiporate she leads how forward-deployed talent is matched, onboarded and shipped to production.

Need the team to make this real?

Describe your need in plain English, get the exact hire, forward-deployed talent or a fractional leader, vetted and matched in 72 hours.

Scope your need →

Keep reading

The Weekly Brief

Intelligence for building AI-native organizations.

One email a week: the sharpest thinking on AI hiring, infrastructure, teams and strategy, for the people building the future of work.

Join operators, founders and CTOs. No spam, unsubscribe anytime.