AI strategy for software companies
An AI strategy for a software company is no longer a question of which model to use. Every competitor buys capability from the same shelf, at the same price, on the same terms. What is left to decide is where your moat moves to, what unit you sell, and which work stops being billable.
Sources cited and dated
No vendor pitch
Updated as the series runs
Get Your Free Consultation!
What an AI strategy for a software company
Every software business now has access to the same frontier models, at roughly the same price, on the same terms. That is a genuinely new condition. The capability that used to separate a strong engineering team from an average one can be bought by both this quarter, with a card, in an afternoon.Most published AI strategy advice answers a different question: how to adopt the tooling. Roadmaps, pilots, budget splits, change management. Useful, but it assumes the business model on the other side is unchanged. It usually is not.This page is our working answer to the harder question underneath: once the tooling is common to everyone, what do you sell, what can you still honestly call difficult to copy, and which parts of today's invoice quietly stop being defensible?
Four questions that decide
Each one is a decision you are already making, whether or not you have made it explicitly. Each maps to one part of the series at the bottom of this page.
Does capability still differentiate you?
01
When the same models are available to you and to the firm you lose deals to, capability stops being a position and becomes a cost line. The useful question is what you were actually selling underneath it — judgement, context, accountability, or just throughput.
Where does the moat move to?
02
Defensibility shifts from what you can do to what you have and what you are trusted with — proprietary data, depth across a whole workflow, and accountability for systems running in production. Prompts and fine-tunes are copied in a fortnight.
What is the unit you sell?
03
Selling a tool puts you in a race against the model. Selling an outcome turns every model improvement into a margin improvement instead of a threat — but it moves delivery risk onto your side of the contract, and only works where the outcome is measurable enough to sign for.
What stops being billable?
04
Some of the work that fills timesheets today will not survive the next capability jump. Deciding which parts — and repricing before a client reprices you — is the whole exercise. Doing it late is how a services business discovers its margin has already gone.
Three things that cannot be bought
If model access is common, defensibility has to come from somewhere else. These are the three that survive contact with a competitor running identical tooling.
Proprietary data
Context that exists only inside your business or your client's — messy, unstandardised, rarely documented, and absent from every training set on the market. It compounds: the longer the system runs, the further ahead it gets.
Workflow depth
Being wired into an entire process rather than one step of it, so replacing you means replacing an operation instead of swapping a feature. Depth is what turns a nice tool into something nobody has the appetite to rip out.
Accountability
Someone answerable when the system is wrong at two in the morning. Models do not carry liability, sign SLAs or attend the post-incident review. Organisations do, and in regulated work that is most of what is being bought.
The pricing shift is
The argument above is ours. These two numbers are not — they are what the analysts currently project, and they are the reason question 03 is urgent rather than theoretical.
What it means
If a meaningful share of budget stops being paid per seat and starts being paid per result, the question is not whether your pricing changes. It is whether you choose the new unit or inherit it.
Figures read 22 September 2026.
The stack that makes a moat
Strategy that never reaches production is a slide. These are the layers where defensibility is actually built, and the tools we use at each one.
Model layer
Interchangeable by design, so that swapping a provider is a config change and not a rewrite.Claude · GPT · Gemini · Llama · Mistral · Amazon Bedrock · Azure OpenAI · Vertex AI
Orchestration and agents
Where a model becomes a workflow: tool use, retrieval, multi-step tasks, human approval gates.LangChain · LlamaIndex · Model Context Protocol · Temporal · Celery
Data layer
The part that is genuinely yours. Pipelines, warehousing and retrieval over context no competitor holds.PostgreSQL · pgvector · Pinecone · Snowflake · BigQuery · dbt · Airflow · Kafka
Evaluation and observability
How you prove the thing works, catch regressions on a model update, and keep a record of why it answered that way.Langfuse · LangSmith · OpenTelemetry · Prometheus · Grafana · custom eval sets
Delivery and infrastructure
Shipping on a schedule you control, with the ability to roll back a bad release in minutes rather than days.Docker · Kubernetes · Terraform · GitHub Actions · AWS · Azure · GCP
Governance and security
The layer that makes accountability sellable: access control, data residency, audit trails, redaction.SSO and RBAC · PII redaction · audit logging · data residency controls
From argument to
The same five steps we run with clients who are working out which parts of their business the next capability jump touches.
The argument is here.
Each question above lands in one of these. Start wherever your own answer was least convincing.
AI and ML development
Model integration, agents, retrieval and evaluation — questions 01 and 02.
Data engineering
The pipelines and warehousing that make proprietary data usable — question 02.
Digital transformation
Rebuilding the operation around the new unit of value — question 03.
DevOps services
Shipping and rolling back fast enough to be accountable — question 04.
MVP development
The thin slice in step 04, built to prove one workflow end to end.
Staff augmentation
Adding the people to run the change without pausing the roadmap.
Book a free consultation
Want help working out which of these applies to you?
Tell us where you are and what you're weighing up. We'll come back with an estimate and a timeline.
Faq
FAQ about AI strategy for software companies
An AI moat is whatever stops a competitor reproducing your product once they have the same models you do. Access to a foundation model is not one, because it is sold to everyone on the same terms. The durable candidates are proprietary data, depth across an entire workflow, and being accountable for a system running in production.