Delivery Plan · Core
Now $4,500£3,750R79,000/ monthWas $8,000£6,700R140,000
Expert project manager
Unlimited backlog
100% your code and IP
Cancel anytime
Start your Delivery Plan
Onboarding starts at checkout · Billed monthly, excl. VAT
Generative AI Development on a Monthly Delivery Plan
Get generative AI features that work in production and use your own data, from LLM products to governed AI agents. Through the Software Factory Delivery Plan, you get one team, one predictable monthly price and one prioritised backlog. Add mobile, web, backend, data, QA or DevOps work to the same backlog whenever you need it.
Delivering since 201248 hours to first codeManaged live on Scrums.com
Use your plan for generative AI development.
14DAYSFirst deliverable in 14 days or your first month is free.
ThousandsOf companies since 2012
Trusted by engineering leaders at ambitious startups, enterprises and global brands.
Read client stories →
How your Software Factory Delivery Plan works
How the generative AI development process works
- 01Queue your GenAI workPick GenAI scopes or describe your own use case.
- 02SubscribeCheck out and onboarding starts immediately.
- 03First code in production in 48 hoursAccess to your repos, model APIs and CI.
- 04Build, evaluate, repeatTrack latency, token cost and hallucination rate on the Scrums.com Platform.
Scrums.com Platform
Track every model, prompt and release gate on the Scrums.com Platform
Goals break down into tasks. Milestones are agreed before work begins. You track every feature, evaluation and release from use case to production, and the next task starts the moment one ships.
Requirement9
Re-ranking layer for policy search
AI5 pts
Ingestion pipeline for the contract archive
Data5 pts
Prompt injection defences on the chat input
AI3 pts
Streaming responses through the API layer
Backend3 pts
Delivering1
Permission-aware retrieval for HR documentsAI
- Prompt and retrieval tests written
- Evaluation set passes in CI
- Code review approved
- Evaluation gate passed
AI engPR #964
Gate
Live38
Citations in every answer
AI · just nowHallucination evaluation set v2
QA · 4d agoToken cost and latency per feature
DevOps · 6d agoModel routing for document summaries
AI · 8d agoIllustrative demo with sample data. On the Scrums.com Platform, the board shows your own goals, tasks and releases.
Scope and capacity
What one workstream delivers, for GenAI or any software
Your backlog can hold any number of requests, but the team delivers them one workstream at a time. A workstream is one managed team working on your highest-priority request until it is live, then starting the next. More requests mean more months, not faster delivery. Work is scoped and delivered within your plan’s capacity. Priorities can change; additional parallel workstreams require more capacity. The same plan covers mobile, web, backend, data, QA and DevOps work, so all of it can share one backlog.
- Active at once
- One request
- To first code in production
- 48 hours
- Sprint cadence
- 2 weeks
- Owned by
- You
Example: one GenAI use case on one workstream
Illustrative- Month 1Discovery and architecture: use cases, compliance constraints, a data audit and model selection, then an architecture decision record. Environment setup follows, and the first production sprint starts.
- Month 2Two-week sprints end in demos of working AI features. Prompt engineering and model evaluation run continuously, with latency, token cost, hallucination rate and user acceptance tracked.
- Month 3Scale and optimise: model performance reviews and fine-tuning iterations from production usage, and corpus expansion as your data grows.
A sample GenAI workstream. It matches the two-week sprint cadence. Onboarding starts immediately after checkout, and your delivery owner scopes your own plan with you.
What affects the number of months
- Use casesThe number of use cases, user journeys and integration points to build.
- Data readinessOutput quality depends on context quality. Thin or unstructured data needs a data foundation first.
- Retrieval or fine-tuningRAG is faster to update. Fine-tuning needs labelled data, an evaluation framework and an MLOps pipeline.
- Data handlingPII redaction, self-hosted models for data sovereignty and audit logging add work up front.
- ComplianceSOC 2, GDPR, POPIA and ISO 27001 alignment is planned from discovery and adds documentation work.
- Review and approvalsHow quickly your team agrees acceptance criteria and signs off releases.
Capabilities
LLM development: products, agents, retrieval, tuning and model operations.
LLM-powered products
Documents · conversation · retrievalLLM-powered products built from scratch: document intelligence systems, conversational interfaces, content generation engines and knowledge retrieval applications. Architecture through deployment, with production hardening included.
When it fitsThe feature must understand, generate or summarise natural language at scale.
Trade-offsNo LLM system eliminates hallucinations entirely. Grounding, output validation and review gates contain them to an acceptable rate for the use case.
How the plan covers it
Queue this scope →One use case taken from idea to an evaluated pilot
Model-agnostic architecture you can evolve
Streaming responses for low-latency interfaces
Frontend components in your existing UI framework
Custom AI agents
Multi-step · governedAutonomous AI agents for multi-step workflows, document processing, compliance monitoring and customer support, governed through the AI Agent Gateway for enterprise-grade control.
When it fitsA workflow waits on human judgement: document classification, contract review, compliance checking or escalation routing.
Trade-offsAn agent acts inside a live workflow, so escalation rules and handoff to people are part of the build.
How the plan covers it
Queue this scope →Agents deployed into a live workflow, not a sandbox
Tool actions, escalation rules and handoff to people
Audit logging for every model decision
Hours saved reported for each rollout
LLM integration and model tuning
API · self-hosted · open-sourceLLM capabilities integrated into existing platforms through an API or a self-hosted deployment. Open-source models such as Llama and Mistral are fine-tuned on your proprietary data where off-the-shelf accuracy falls short.
When it fitsYou cannot rebuild the core product: an LLM integration layer sits on top of your existing stack.
Trade-offsFine-tuning needs labelled training data, an evaluation framework and an MLOps pipeline. RAG is often faster to update and easier to audit.
How the plan covers it
Queue this scope →REST and GraphQL layers into existing application logic
Fine-tuning for domain language, tone and task accuracy
Requests routed to the model best suited to each function
Token cost and latency visible per feature
Grounded answers over your own knowledge
Pinecone · Weaviate · pgvectorRetrieval-augmented generation (RAG) connects an LLM to your proprietary knowledge base for grounded, verifiable responses: vector database infrastructure, ingestion pipelines and re-ranking layers, with your data kept out of third-party training.
When it fitsYour proprietary data would make a generic LLM far more useful, and answers must be verifiable against a known source.
Trade-offsAnswer quality depends on context quality. Thin or unstructured data needs a data foundation first.
How the plan covers it
Queue this scope →Ingestion pipelines over documents, policies and product content
Re-ranking and citation so every answer is checkable
Permission-aware retrieval for internal knowledge
Corpus refresh as your content grows
GenAI workflow automation
Contracts · regulation · reportsWorkflows that need natural language understanding, automated: contract analysis, regulatory document review, customer communication drafting, report generation and knowledge base maintenance.
When it fitsContract analysis, regulatory document review or report drafting waits on people to read and write.
Trade-offsPeople still review the exceptions and the high-stakes outputs.
How the plan covers it
Queue this scope →Structured data extracted from unstructured documents
Validation and exception handling on every run
People review only the exceptions and high-stakes outputs
Event-driven triggers through webhooks
LLM optimisation and ongoing maintenance
SLA-backed model operationsOngoing model monitoring, prompt optimisation, RAG corpus refresh and version migration as providers release updates. Token costs and latency are tracked as first-class delivery metrics.
When it fitsA GenAI system is live, and new model versions, prompt drift or stale content put its quality at risk.
Trade-offsMaintenance is ongoing work, not a one-off project. It shares the backlog with new features.
How the plan covers it
Queue this scope →Model version evaluation and migration planning
Prompt optimisation from production usage patterns
Retrieval quality monitoring and corpus refresh
Cost optimisation as usage scales
What's included in generative AI development
Architecture, retrieval, agents and model operations. Managed together.
Every Delivery Plan · Core comes with the full GenAI team and the tooling around it: architecture, LLM engineering, retrieval, evaluation and model operations, run by one delivery owner.
Delivery managementOne accountable delivery owner.Milestones agreed before work begins.
01Discovery and architectureUse cases, a data audit and model selection, then an architecture decision record and a sprint plan
02LLM integrationREST and GraphQL layers, event-driven triggers and streaming responses into your existing UI
03Retrieval (RAG)Pinecone, Weaviate and pgvector, with ingestion pipelines and re-ranking layers
04Fine-tuningOpen-source models such as Llama and Mistral tuned on your proprietary data
05Custom AI agentsMulti-step agents governed through the AI Agent Gateway, with audit logging
06Hallucination controlsGrounding, output validation, confidence scoring and human review gates
07Security and compliancePII kept out of model calls, prompt injection defences, SOC 2, GDPR, POPIA and ISO 27001
08Model operationsVersion migration, prompt optimisation, corpus refresh and cost optimisation
10Code and IP ownershipAll code, pipelines and IP are yours from day one.
100%Launch is not the finish line.
Model version migrationNew model versions evaluated, and migrations planned before breaking changes land.
Prompt optimisationPrompts tuned from production usage as user behaviour changes.
Retrieval refreshRetrieval quality monitored and the corpus refreshed as content grows stale.
Cost optimisationToken costs benchmarked against your quality thresholds as usage scales.
Runs on the Scrums.com platform
Every Delivery Plan runs on the Enterprise AI Platform for software engineering.
Your team, its AI agents and your delivery data sit on one governed platform. That is what powers your delivery board, the AI Agent Gateway and the talent bench behind your plan.
Your planDelivery Plan · CoreAI, data, backend, web, mobile and QA scopes on one backlogSee pricing →TalentAI engineers and teamsCertified engineers, CTOs and dedicated teamsHire engineers →AI Catalog163 agents · 101 MCPs · 30 modelsDeployable AI for engineering orgsBrowse the catalog →PlatformEnterprise AI PlatformAgent harness, AI and MCP gateways, governanceExplore the platform →
What your GenAI team draws on, already part of your planBrowse the full AI catalog →
Engineering talent
AI engineers and teams
AI engineers with production experience across OpenAI, Gemini, Llama, Mistral, Claude and Grok, from the Scrums.com bench, matched to your use case and domain.
AI Catalog
Agents, MCPs and models for GenAI
Deployable agents, models and MCP servers live in the AI Catalog. The AI Agent Gateway governs the agents you build and the ones you deploy.
AI infrastructure
Infrastructure for models and retrieval
Data and AI infrastructure for your models and retrieval, deployable from the catalog with your $500 first-month credits.
Compare
One AI hire covers one role. The plan covers the whole team.
Production GenAI needs architecture, LLM engineering, retrieval, evaluation and model operations. Hiring covers them one role at a time. The Delivery Plan covers them in one monthly price.
Compare
RecommendedScrums.com Delivery Plan · Core
In-house AI engineer
Traditional AI agency
Roles covered
Architecture, LLM engineering, retrieval, evaluation, model operations
One role per hire
Depends on the agency bench
Time to start
48 hours to first code in production
A hiring cycle first
Proposal and contract cycle first
How you pay
Salary, benefits and recruiting fees
Project quotes and change requests
Who runs delivery
One accountable delivery owner
You manage the work
The agency project manager
Visibility
Latency, token cost and hallucination rate on the Scrums.com Platform
Your own tooling
Status reports from the agency
40-60%†lower cost than US and UK firms
Start for Basis: the figure that the current Scrums.com generative AI development page states against US and UK firms. It is not a quote for your scope.
What you get
Why companies choose managed generative AI consulting services
01
Stays in production
The engineers who build your system stay across model updates, prompt drift and stale content.
02
One monthly price
One plan price for architecture, build and model operations. Your scope sets the number of months.
03
Model-agnostic
Model selection follows your use case, latency, cost targets and data residency, not vendor agreements.
04
Compliance built in
Data residency, PII in prompt context, audit trails and output governance are planned from discovery.
05
Integrated with your stack
LLM capabilities land inside your existing platforms, APIs and workflows. No platform rewrite is required.
06
Quality you can measure
Hallucination rate, latency and token cost are tracked from sprint one.
Customers
Generative AI development case studies
Teams are matched to your use case and domain, with regulated delivery for FinTech, banking and insurance, across OpenAI, Gemini, Llama, Mistral, Claude and Grok.
Featured GenAI case study
How AI cut an 8-year project to just 2 years
- Challenge
- A client faced an 8-year custom ERP build or multi-million-dollar SAP or Oracle licensing. The legacy code was in a non-English language, and discovery and documentation were manual.
- Platforms and technologies
- AI-driven software engineering on SEOP and the AI Agent Gateway.
- Work delivered
- AI-generated ERDs, SRS documents and Jira tickets, AI code generation for 90% of boilerplate and business logic, and continuous sprint velocity tracking.
- Measurable result
- Project timeline from 8 years to 2 years. Discovery from 5–6 hours per module to 1 hour across 10 modules. 56 modules delivered in the timeframe that delivered 48 before.
Scopes
Generative AI development scopes you can queue
One Delivery Plan · Core covers your development priorities across services. Start with GenAI, then queue mobile, web, backend, data or QA scopes on the same backlog, or describe your own.
GAIGenerative AI10 scopesAIAI developmentExplore ↗DEData engineeringExplore ↗APIBackend & APIsExplore ↗WWeb appsExplore ↗MMobile appsExplore ↗QAQA & testingExplore ↗OPSDevOps & reliabilityExplore ↗→All scopesCatalog
SprintSC-01
AI Data Infrastructure Foundation
Set up the data pipelines, versioning, feature flow, and storage patterns required for repeatable ML and AI work. Finish with an ML-ready foundation running in your cloud.
SprintSC-02
AI Feature Productionization
Take an existing AI proof-of-concept into monitored production with reliability, evaluation, security, and cost controls. Finish state: the demo that impressed everyone, now boring, monitored, and live.
ScopeSC-03
AI Ideation Workshop
Facilitated 2-week workshops that align stakeholders, define goals, and produce an actionable design & development strategy.
SprintSC-04
AI Model Routing & Cost Optimization
Route requests across models based on quality, latency, availability, and cost while tracking spend and outcomes. Finish state: the right model per request, spend visible per feature, quality protected by evaluation.
SprintSC-05
Document Extraction AI Workflow
Extract structured fields and classifications from documents with validation and exception handling. Finish state: documents becoming clean structured data automatically, with humans only on exceptions.
SprintSC-06
Enterprise Knowledge Chatbot
Deploy a permission-aware conversational interface with RAG ingestion, retrieval, and grounding over approved internal knowledge. Finish state: staff getting cited, permission-correct answers from your own content.
SprintSC-07
GenAI Pilot Build
One LLM use case taken from idea to a production pilot: use-case and data audit, model selection, RAG and integration build, and an evaluated pilot running against real usage — engineering, not a demo.
SprintSC-08
In-App AI Copilot
Embed a scoped AI assistant inside an existing product with context, permissions, and basic evaluation. Finish state: a useful copilot live for real users, doing a defined set of things well.
SprintSC-09
LLM Evaluation & Guardrails Package
Create test sets, scoring, red-team checks, policy controls, and release gates for an LLM-powered feature. Finish state: AI quality and safety measured and gated, not vibes-checked.
Managed SLASC-10
Model Ops Retainer
SLA-backed operations for production ML and LLM systems: model and prompt monitoring, evaluation and drift detection, RAG corpus refresh, and managed version migrations — because the failure mode of production AI is abandonment.
The plan in full
Generative AI development services pricing
Core costs Was $8,000£6,700R140,000 now $4,500£3,750R79,000/month for one active workstream. Your use case’s scope determines the delivery schedule and total cost. You can use the same subscription for other development priorities.
Best for companies whose product must understand or generate language, whose own data could ground a generic LLM, or whose workflows wait on human judgement.
Delivery Plan · Core
One plan for your software backlog.
Now $4,500£3,750R79,000/ monthWas $8,000£6,700R140,000
One active workstream. Managed delivery.
Start your Delivery Plan →
Billed monthly · excl. VAT/sales tax · onboarding starts immediately after checkout
Included every month
Delivery owner. One accountable owner runs your AI engineers, data engineers and QA engineers.
One backlog across your development priorities. You set the order; requests are scoped and delivered one workstream at a time.
Visibility. Milestones agreed up front. Latency, token cost and hallucination rate live on the Scrums.com Platform.
Ownership. Code, pipelines and data are yours from day one, with handover included.
Work is scoped and delivered within your plan’s capacity. Priorities can change; additional parallel workstreams require more capacity.
First-month benefits
14DAYS
First-delivery guaranteeYour first agreed deliverable ships within 14 days of scope and access being confirmed, or your first month is free.
$500 in AI infrastructure catalog creditsMONTH 1
Annual-plan benefits$10,000 in specialist services
Technical Assessment & Delivery RoadmapWe review your systems, rank your priorities and agree the first deliverable.
CI/CD & Quality Pipeline SetupYour pipelines, environments and quality checks, set up and handed over to you.
You buy monthly delivery capacity, not a fixed project price. Infrastructure usage and third-party services are billed separately. Full Delivery Plan coverage and terms →
Need more parallel workstreams? Talk to us →
Day 0Scope and access confirmedOnboarding starts at checkout. You confirm scope and share repos, model API access and CI.
Within 48 hoursFirst code in productionFirst change merged and deployed through your pipeline.
Within 14 daysFirst agreed deliverable shipsFor example the architecture decision record and sprint plan for your first use case.
If it does not shipYour first month is freeYou do not pay for month one.
The first deliverable is agreed in writing before the clock starts. The clock starts when scope and access are confirmed. Applies to the first deliverable on a new plan.Read the full Delivery Plan terms →
FAQ
Generative AI development questions, answered
Still deciding? Talk to sales and get a scoped answer for your use case.
Delivery Plan · Core costs Was $8,000£6,700R140,000 now $4,500£3,750R79,000 a month, excluding VAT or sales tax. That buys one active workstream with a managed GenAI team. It is monthly delivery capacity, not a fixed price for a complete AI product, so your scope sets the number of months.
Yes. Delivery Plan · Core covers mobile, web, backend, data, QA and DevOps work as well as generative AI development. All of it sits in one prioritised backlog and is delivered within your plan’s capacity, one active workstream at a time. Priorities can change; additional parallel workstreams require more capacity.
Discovery and architecture, LLM integration, retrieval-augmented generation (RAG), fine-tuning, custom AI agents, hallucination controls, security and compliance, model operations, and one accountable delivery owner. Every plan includes $500 in AI infrastructure catalog credits in your first month. Annual plans also include $10,000 in specialist services: a technical assessment, a delivery roadmap and CI/CD pipeline setup.
You rank the backlog. Your delivery owner scopes the top request, agrees acceptance criteria with you, and the team delivers it. You can reorder anything that has not started. One request is active at a time on Core; additional parallel workstreams require more capacity.
The clock starts when you confirm scope and give access to your repositories, model APIs and CI. Within 48 hours the team has its first code in production in your environment. Within 14 days your first agreed deliverable ships, for example the architecture decision record and sprint plan for your first use case. If it does not, your first month is free.
48 hours to first code in production means the first change merged and deployed through your pipeline, counted from the day scope and access are confirmed. 14 days means the deliverable named at kickoff ships within 14 days of that date, or your first month is free. “Thousands of companies” counts organisations that have bought engineering services from Scrums.com since 2012. A two-week sprint cadence and six model families in production are the figures the current Scrums.com generative AI development page states. 40-60% lower cost is the figure that page states against US and UK firms; it is not a quote for your scope.
Generative AI development is the engineering discipline of building products and systems powered by large language models and other generative AI technologies. It covers LLM integration, custom AI agent development, retrieval-augmented generation (RAG) systems, fine-tuning proprietary models and deploying generative AI features into production applications.
Scrums.com is model-agnostic. Engineers have production experience with OpenAI (GPT-4o, o1, o3), Google Gemini, Meta Llama (open-source, self-hosted), Mistral, Anthropic Claude and Grok. Model selection follows your use case, latency requirements, cost targets and data residency constraints, not vendor agreements. Where appropriate, multiple models work in combination, each routed to the task it suits best.
Retrieval-augmented generation (RAG) connects an LLM to a searchable knowledge base of your own content, so the model grounds its responses in your proprietary data without exposing that data to third-party training. RAG is the recommendation when you have a large corpus of internal documents, policies or product content, when responses must be verifiable against a known source, or when hallucinations are unacceptable. RAG often comes before fine-tuning, because it is faster to update and easier to audit.
Yes. Open-source models such as Llama and Mistral are fine-tuned on your datasets for domain adaptation, tone consistency and task accuracy. Fine-tuning needs sufficient labelled training data, a well-defined evaluation framework and an MLOps pipeline for model versioning, deployment and performance monitoring. The team advises whether fine-tuning or RAG suits your use case before you commit to either.
Hallucination prevention is an engineering problem, not only a prompt engineering problem. The approach combines RAG to ground responses in verified sources, output validation against known facts or structured constraints, confidence scoring and fallback handling, human review gates for high-stakes outputs, and continuous monitoring of output quality as a delivery metric. No LLM system eliminates hallucinations entirely; well-engineered systems contain them to acceptable rates for their use case.
Data handling architecture is designed before any code is written. For regulated clients that means keeping sensitive PII out of third-party model API calls by pre-processing or redacting inputs, self-hosted open-source models where data sovereignty requires it, role-based access controls on AI outputs, full audit logs of model inputs, outputs and decisions, and alignment with SOC 2, GDPR, POPIA and ISO 27001 where applicable.
Providers release new model versions, sometimes with breaking changes. Prompt performance drifts as user behaviour changes, retrieval quality degrades as content grows stale, and token costs change as usage scales. Maintenance covers model version evaluation and migration, prompt optimisation from production usage, corpus refresh and retrieval monitoring, and cost optimisation against your quality thresholds.
First code reaches production within 48 hours of confirmed scope and access. The first sprint covers discovery, architecture, environment setup, model selection and kickoff. After that, two-week sprints end in demos of working AI features.
GENERATIVE AI DEVELOPMENT · SORTED
Ship GenAI to production, then keep it there.
Delivery Plan · Core: one monthly price, one backlog across your development priorities, one active workstream and a managed team for LLM products, retrieval, agents and model operations.
LLM PRODUCTS · AGENTS · RETRIEVAL · MODEL OPS