Everyone says “you must invest in AI in 2026” – but no one wants to admit how many AI licenses are already sitting unused.
At Pynest, we build software and augment engineering teams for clients in the US and Europe. That means we sit on both sides of the table: as a buyer of tools like Microsoft 365 Copilot, and as a partner who has to make sure these tools really fit into daily engineering and business workflows.
This is how we actually decide when AI software is worth paying for – and when it is smarter to wait.
The context: AI budgets are under pressure
Analysts are already warning that the first wave of “buy AI everywhere” is cooling down. A recent Forrester view suggests enterprises are postponing a significant share of planned AI budgets because returns are not matching promises.
At the same time, research summarized by Microsoft and IDC shows that where generative AI is integrated deeply into operations, early adopters can see strong ROI, not just nice demos.
In other words: AI software is not automatically good or bad. It depends entirely on integration, use case selection, and how people actually work with it.
Our basic rule: start from workflows, not from logos
We do not begin with the question “Should we buy Copilot or Agentforce?”. We start with “Which workflows are currently painful or expensive?”
Typical examples inside Pynest:
- Drafting and refining internal docs, RFCs, and client emails.
- Sifting through long project threads and meeting notes to understand “what really happened”.
- Helping engineers explore a new codebase faster, without turning them into prompt operators.
Only after we identify the top 3–5 painful workflows do we map them to tools. Sometimes an AI feature inside software we already pay for is enough; sometimes we need a dedicated product; and often we decide to build a small internal agent instead of buying a big platform.
For us, buying AI is never “strategic” in the abstract. It must be brutally tactical.
A simple evaluation framework we actually use
When we look at tools like Copilot-style assistants or AI-enhanced CRMs, we use three lenses:
Value per active user, not per seat.
We estimate how many hours a typical engineer, recruiter, or manager would realistically save per week. If the annual license cost is higher than 20–30% of that time value, we are already cautious.Integration friction.
Does the tool plug into our existing stack (Git, ticketing, docs, HRM) or will it create yet another silo? If it cannot see the same context our people work in, we know adoption will stall.Data and risk profile.
We check where prompts and outputs are stored, what training the vendor does on our data, and how they handle access control and audit trails. If this is unclear, the deal pauses – even if the UX is beautiful.
We treat vendor promises with healthy skepticism. A Microsoft-sponsored IDC report may show strong ROI averages, but we always translate that into our own numbers and workflows before committing.
What other experts are seeing
We are not alone in this more cautious approach.
Forrester VP research director Emily Collins notes that 2025 for CMOs is “less sensationalized and more operationalized”: AI investments must prove real efficiency and effectiveness, not just generate headlines.
Data scientist and author Daniel Gutierrez has highlighted the same IDC study: generative AI delivers substantial ROI only when embedded into core operations, not as a side experiment.
KPMG’s recent “Future of Work” research also points out that companies get far greater financial benefit from AI when employees themselves feel it helps them in daily work – not when it is pushed only from the top.
These outside views match what we see on the ground: tools succeed when they sit in the flow of work and are trusted by the people using them.
How we decide to adopt, delay, or avoid an AI tool
When we adopt
We tend to adopt when all of the following are true:
- There is a clear, repeatable workflow with measurable pain (for example, engineers writing repetitive test boilerplate or recruiters cleaning up CVs).
- The tool integrates with our stack with minimal extra steps.
- We can define simple, concrete success metrics within 3–6 months (time to complete a task, ticket volume, lead time, etc.).
- At least one “product owner” inside the business is personally invested in making it work.
For example, before buying a Copilot-type license at scale, we run a small internal pilot with detailed measurement: which teams use it, how often, and what changes in their commit patterns, defect rates, or lead times.
When we delay
We delay when:
- The business case is vague: “everyone in the market is buying this, we’ll figure it out later.”
- The vendor’s data policy is unclear or evolving too fast.
- Our people do not yet have basic AI literacy – they need training first, not more tools.
Given that some reports show a significant share of enterprises are postponing AI budgets to later years because value is unclear, we are comfortable saying “not now” even if a tool is trendy.
When we avoid completely
We avoid tools that:
- Lock our data in a proprietary format with no realistic export path.
- Require sending sensitive client code or personal data to a black-box environment.
- Promise to “replace entire teams” instead of augmenting them.
In our experience, anything that starts with “this will replace your engineers / analysts / SDRs” almost never delivers – and almost always creates cultural damage.
Budgeting and ROI: how we talk about numbers
On the budgeting side, we treat AI tools like any other operational investment:
- We cap AI licenses as a share of our total SaaS spend, so they do not silently swallow the entire budget.
- We tie renewal to observed behavior, not to vendor roadmaps: if usage and impact drop, we downgrade or cancel, even if the vendor is promising amazing new features “next quarter.”
- We measure “time to useful output”, not just raw hours saved. If a tool makes a developer faster but increases review time and error rates, we count that as a net negative.
External stats are useful as a sanity check: for instance, some analyses show that many businesses now allocate at least 5% of their digital budget to generative AI and report measurable productivity gains – but these are averages, not a prescription.
What this looks like inside Pynest
Concretely, in 2025–2026 our approach at Pynest looks like this:
- We run small, focused AI pilots in real teams (engineering, HR, sales ops) with clear owners and metrics.
- We prefer platforms that extend tools we already use (IDE, office suite, CRM) over adding more standalone products.
- We invest heavily in people and process: training, internal champions, and guidelines on when not to trust AI output.
- We regularly review all paid AI tools twice a year and are not afraid to cut what does not prove its value.
For us, AI software is not a magic line item. It is one more tool in the toolbox. The only real question is: does it make our people, and our clients’ teams, better at the work that actually moves the business forward?
If the honest answer is “yes, and we can prove it,” then we sign the contract. If not, we wait.
Top comments (0)