LLM Engineering

How long does a custom LLM project take?

Modulus May 15, 2026 6 min read

The most common question buyers ask before signing an LLM engagement is not "how much does it cost?" — it is "how long will it take?" The answer is never a clean number, but it is rarely a mystery either. Timelines are determined by a small set of knowable factors, and once you understand them, you can scope a project with a reasonable confidence interval before you write a single check.

This piece breaks down realistic delivery windows for the most common LLM project types, what inflates timelines in practice, and the questions you should ask any vendor before you commit to a start date.

TL;DR
  • A prompt-engineering / API integration project ships in 2–6 weeks.
  • A production RAG system takes 6–14 weeks end-to-end, including data preparation.
  • Fine-tuning a domain-specific model adds 4–10 weeks on top of the base build.
  • Full custom LLM development (pretraining or continued pretraining) is a 4–12 month effort.
  • Data readiness and integration complexity are the two primary timeline killers — not model training itself.

The project type determines the ceiling, not the floor

Before you can estimate time, you need to agree on what you are actually building. The term "custom LLM project" covers a spectrum so wide that two projects with the same label can differ by six months in delivery time. The four canonical project types — and their honest timelines — are:

Project type Typical delivery window Primary driver of variance
Prompt engineering & API integration 2–6 weeks Stakeholder alignment, API access approval
RAG pipeline (production-grade) 6–14 weeks Data quality and volume, retrieval architecture choices
Fine-tuning on domain data 8–16 weeks Training data curation, evaluation harness setup
Custom pretraining / continued pretraining 4–12 months Compute procurement, data pipeline engineering, safety evaluation

These are not pessimistic estimates. They reflect reality when you include data preparation, evaluation, security review, and deployment — the phases that agencies and vendors reliably omit from early estimates. Custom LLM development done properly has a testing and hardening phase that is routinely 30–40% of total project time.

Phase 1: Discovery and scoping (1–3 weeks)

The scoping phase is the one buyers consistently underestimate. A competent vendor will spend one to three weeks doing nothing but understanding your use case, your data, your existing infrastructure, and your success criteria. This is not billable padding — it is the only way to produce a timeline estimate you can trust.

During scoping, the vendor should be answering: What does success look like quantifiably? What does your training or knowledge data look like, and how clean is it? What systems does the LLM need to connect to? Who needs to approve security and compliance requirements? Every one of these questions that goes unanswered in week one adds weeks downstream.

The cheapest thing you can do before engaging a vendor is to prepare a data inventory. Know what you have, where it lives, what format it is in, and who controls access to it. Vendors who get a clean data inventory on day one ship faster than those who spend week three still waiting for credentials to an internal document store.

Phase 2: Data preparation (2–8 weeks, often parallel)

This is the phase most buyers are surprised by. Whether you are building RAG, fine-tuning, or a full custom model, data preparation is almost always the longest single phase and the one most subject to scope creep.

For a RAG system, data prep means: cleaning and normalizing source documents, structuring metadata, building ingestion pipelines, handling edge cases (scanned PDFs, proprietary formats, legacy databases), and establishing a process for keeping the knowledge base current. For fine-tuning, it means curating high-quality input-output pairs — a task that is deceptively labor-intensive when done correctly.

The most common timeline disaster in LLM projects is discovering mid-project that the source data is worse than expected. Inconsistent terminology, stale records, version conflicts, missing context — all of these require remediation before a model can reliably use the data. Budget 30–50% of your project timeline for data work and you will be close to reality.

Phase 3: Model development and integration (3–8 weeks)

Assuming clean data and a stable architecture decision, the core development phase for most projects falls in the three-to-eight-week range. For a RAG pipeline, this covers embedding generation, vector store setup, retrieval chain logic, generation prompt engineering, and API layer construction. For fine-tuning, it covers training runs, hyperparameter tuning, and initial evaluation passes.

Integration with your existing systems is often where this phase extends beyond expectations. Connecting an LLM to a CRM, ERP, or proprietary database involves authentication, rate limiting, data transformation, and edge-case handling for every data source. Projects with three or more integration points should add two to four weeks to baseline estimates.

See our companion piece on fine-tuning vs RAG cost decision tree for guidance on which architecture to choose before you enter this phase — changing architecture mid-development resets the clock entirely.

Phase 4: Evaluation and testing (2–4 weeks)

Production LLMs require structured evaluation before deployment. This is not optional, and vendors who skip it are setting you up for a painful post-launch patch cycle. A proper evaluation phase covers: automated benchmarks against your specific task distribution, adversarial testing for prompt injection and jailbreak vectors, output consistency testing across edge cases, latency and throughput profiling under production load, and human review of sampled outputs.

For regulated industries — finance, healthcare, legal — add a compliance review layer that can add two to six weeks on its own. Plan for this before you start, not after the model is built.

Our article on building an LLM evaluation harness covers the specific components you need before you can ship with confidence.

Phase 5: Deployment and hardening (1–3 weeks)

Deployment is shorter than most phases but has more potential for last-minute friction. Infrastructure provisioning, secrets management, CI/CD pipeline setup, monitoring and alerting configuration, rollback procedures — each of these has dependencies on your internal DevOps processes that a vendor cannot fully control. The fastest deployments happen when the client has a staging environment ready before the model build is complete.

Post-deployment hardening — the first two weeks of live traffic observation, anomaly response, and prompt refinement — should be scoped explicitly. A vendor who hands over code and disappears is not a production partner; they are a liability.

What consistently makes projects run long

After running custom LLM engagements across multiple industries, the same factors appear in every delayed project:

  • Data access delays: waiting on IT provisioning, legal review of data-sharing agreements, or security approval for external tooling.
  • Undefined success criteria: when no one can agree on what "good enough" looks like, evaluation loops become infinite.
  • Scope creep in integrations: "while you're in there, can we also connect it to X" mid-project.
  • Stakeholder availability: subject-matter experts who are needed for data labeling or output review but have no calendar capacity.
  • Compliance reviews that were not scoped: particularly common in finance and healthcare verticals.
  • Architecture pivots: choosing RAG then switching to fine-tuning (or vice versa) after development has started.
  • Infrastructure procurement: cloud GPU access or on-premise hardware that takes weeks longer than expected to provision.
  • Underspecified vendor contracts: deliverables that are ambiguous enough to require renegotiation mid-engagement.

Questions to ask your vendor before signing

A vendor who gives you a confident timeline in a first meeting without asking about your data, integrations, or compliance requirements is not trustworthy. The right questions to press any custom LLM development vendor on are:

  • What is your process for data quality assessment, and what happens if we discover data issues after the project starts?
  • Which phases are fixed-scope and which are time-and-materials?
  • What are the explicit dependencies on our team — approvals, data access, subject-matter expert time?
  • What does the evaluation phase look like, and who defines the acceptance criteria?
  • What is included in the post-deployment period, and what is out of scope?
  • How have your last three projects of this type compared to their initial timeline estimates?

The last question is the most revealing. A vendor who can answer it honestly — including the projects that ran over and why — has the operational maturity to give you a realistic estimate. See our guide on evaluating LLM development vendors for a full RFP template and scoring rubric.

A practical planning framework

If you are building your internal business case or setting expectations with a leadership team, use these planning brackets:

For a well-scoped RAG implementation with reasonably clean data and two to three integrations: 10–14 weeks from contract to production. For a fine-tuned domain model with a proper evaluation harness: 16–22 weeks. For anything involving custom pretraining or a model from scratch: do not promise a quarter — plan in half-year increments.

The most expensive timeline mistake is overpromising to leadership on an LLM project scope. The second-most expensive is choosing a vendor based on who gives the shortest estimate rather than who has the most honest track record. Both mistakes result in the same outcome: a project that starts fast, hits a wall at week six, and finishes six months late.

Visit our insights library for more practitioner-level guides, or explore our full LLM engineering services to understand how we structure engagements to ship on the timeline we quote.

Category LLM Engineering
← All insights
Related

Know your timeline before you commit.

Fixed-price proposal with a realistic delivery schedule. Free discovery call.