AI in PM, Scaled Agile & Atlassian Williams: PM Highlights, February 2025

February 2025 was a clarifying month for project management. AI tools moved from “coming soon” to “you are paying for them now”. Scaled Agile crossed the threshold from trend to default. The PMI absorbed the Agile Alliance and started rewriting certifications. And Atlassian put its name on a Formula 1 car to prove that integrated tooling is now an enterprise category, not a Jira upsell.

If you are a CTO, founder, or product leader choosing or running a software development partner in 2026, the question is not “is AI ready?”. It is which AI to adopt, what to refuse, and what to demand from a vendor pitching themselves as “AI-powered”. This playbook walks through every February 2025 PM headline, what survived in 2026, and how we run the same decisions inside Fora Soft.

Why Fora Soft wrote this playbook

Fora Soft has been delivering custom software since 2005, with a specialism in video, audio, AI, and real-time communication products. We currently run mixed Scrum/Kanban delivery for clients across EdTech, healthcare, fitness, media, and B2B SaaS. We use Agent Engineering internally, which compresses delivery time on most workstreams by 30–40% versus a baseline team — the data and methodology are documented in our AI-driven development case study.

Recent reference points include BrainCert (an EdTech and virtual classroom platform we have evolved across multiple major releases), Scholarly (a learning platform with 15,000+ users and an AWS Innovation Award), and AppyBee (a fitness booking platform live in 800+ studios across iOS and Android). Our internal PM stack is documented in detail across our project planning, product development, and product launch playbooks — the same processes we used to evaluate every change in this digest.

So when we recommend treating AI estimation tools with caution, ignoring half of the meeting-bot category, or budgeting Copilot like a senior hire, that opinion comes from running these tools across real production teams — not from a vendor demo.

AI in PM — the productivity paradox you cannot ignore

The single most important PM finding from early 2025 is that AI tools do not automatically deliver the productivity uplift their vendors claim. METR’s randomised controlled trial with sixteen experienced open-source developers found that AI tooling actually increased completion time by ~19% — while the same developers believed AI had made them roughly 24% faster.

The opposite case exists too. Other longitudinal datasets across 300 engineers reported cycle time falling roughly a third and review time falling around 30% — with one nasty side effect: pull-request review time grew almost two-fold. The bottleneck simply moved from typing to approving. Both findings are real. They are not contradictory; they describe different teams shipping different work.

Gartner’s 2025 enterprise survey is the sober anchor: only 28% of AI use cases in IT operations fully meet ROI expectations, 20% of pilots fail outright, and 72% of CIOs report breaking even or losing money on AI investments. The teams winning are the ones using AI for narrow, well-defined work — not the ones replacing PM judgement with a chat window.

Microsoft Copilot for PM — what it actually does, and what it costs

Microsoft retired Project for the Web and consolidated PM into Microsoft Planner with Copilot integration. The Copilot for Microsoft 365 add-on lists at $30 USD per user per month on top of a base Microsoft 365 license (Apps for Enterprise, Business Basic/Standard/Premium, or E3/E5/F1/F3).

What you are buying for that price:

1. Risk assessment. Copilot scans project metadata (scope, schedule, budget) and surfaces likely risks plus mitigation suggestions.

2. Task plan generation. AI-suggested task breakdowns, durations, and effort estimates, anchored to historical patterns from similar projects.

3. Natural-language Q&A. Chat interface to query project state, generate status reports, draft updates.

4. Project Manager Agent (preview). Automates task creation, plan generation, and progress tracking with reduced manual input.

For a 20-person engineering team, $30/user adds about $7,200/year — before you measure a single hour saved. The honest question to ask is whether it replaces PM hours, augments them, or just adds another notification stream.

Agile in 2025 — what scaled, what stayed niche

Two things finally settled in early 2025. First, scaled Agile is no longer a debate — SAFe adoption climbed from roughly 37% in 2021 to about 53% by 2025, making it the default enterprise scaling framework. Second, “Agile + Design Thinking” moved from buzzword to standard practice in serious product orgs.

For the projects we run, single-team Scrum with strict 2-week sprints and a Kanban overlay for incident work is still the most reliable setup for product builds up to ~25 engineers. We only reach for SAFe-style scaling when a single product touches 4+ teams with hard cross-team dependencies.

Agile + Design Thinking — problem-finding plus problem-solving

Design Thinking is how you find the right problem; Agile is how you ship the solution. The integration is concrete: a 5-day design sprint (research, ideation, prototyping, testing) feeds into 2-week Agile build sprints, so you stop spending engineering hours on the wrong feature.

If your dev partner skips this step, you will pay later for the pivot. Demand at least one validated design sprint per major feature epic, with a written prototype-test report.

Atlassian + Williams Racing — the system-of-work proof point

Atlassian became title sponsor of Williams Racing in 2025; the team rebranded to Atlassian Williams Racing for the 2025 Formula 1 season, with Atlassian branding on the FW47. The interesting part is not the deal value — it is what got deployed.

Jira for issue and parts dependency tracking across hundreds of suppliers. Confluence as the institutional memory of track conditions, race learnings, and post-mortems. Loom for asynchronous video to replace a chunk of synchronous status meetings. Rovo (Atlassian’s AI assistant) answering routine project-state questions in plain language across all of the above.

Williams scored 111 points in the 2025 season, the team’s strongest result since 2016. Tooling alone did not earn the points, but it removed the kind of friction that historically slowed engineering decisions. The lesson generalises: a single source of truth, AI-assisted Q&A, and async video replacing synchronous standups is now a credible enterprise pattern, not a startup affectation.

PMI × Agile Alliance merger — what changed and why it matters

The Agile Alliance officially joined PMI on December 31, 2024, forming what is now PMI Agile Alliance. A new PMI-ACP exam launches globally on November 8, 2026, with updated certification standards.

Vasco Duarte and other long-time Agile voices have publicly worried that institutional consolidation risks diluting Agile’s emphasis on iteration, feedback, and flexibility. The risk is real but not new — PMP-style waterfall thinking has always crept into “Agile” rollouts. The practical implication is that a fresh PMI-ACP badge in 2027 will be cheaper signal than a track record of running successful retros.

When you evaluate certifications from a partner’s PMs, look past the badge. Ask for two specific things: the most recent retrospective they ran (what did the team change as a result?) and an example of an estimate they had to revise mid-sprint (what did they learn about the work?). The answers tell you whether the certification is decorative or operational.

AI meeting bots — the uncanny valley and the data trade-off

Otter.ai, Read.ai, Microsoft Copilot, and Zoom AI Companion are now mainstream. Otter typically runs 85–90% transcription accuracy and degrades on technical jargon, accents, and rapid speaker changes. Read.ai is SOC 2 Type 2 and HIPAA compliant and adds sentiment analysis and a meeting score. Zoom AI Companion is integrated natively, with claimed zero retention by third-party AI providers.

Two things hold them back. The uncanny valley is real — near-perfect transcription with one wrong attribution feels worse than a fully robotic transcript. And the security trade-off is non-trivial: every meeting bot ships your conversation through a third-party LLM, which matters if you operate under HIPAA, SOC 2, or EU data residency rules.

AI estimation — where it earns its keep, where it lies confidently

AI estimation tools are excellent at bounded, repetitive work and dangerous on novel work. The hallucination rate across LLM outputs is small in absolute terms (often quoted under 2%), but in estimation it is concentrated exactly where you cannot afford it — new integrations, untested architectures, anything without a historical analogue.

Where AI estimation works. Bug fixes, refactors, code-generation sprints with clear scope, repeated CRUD-style features. Anything where the model has a plausible historical baseline.

Where it fails. First-of-kind integrations, research-heavy features, architectural decisions touching multiple systems, anything with human-factors uncertainty. Here AI will produce a confident, plausible, and often wildly wrong estimate.

Our standard practice is bottom-up AI estimation for known tasks plus mandatory senior-engineer uncertainty bounds on anything novel. We document the full approach in our software estimation guide and the data behind our Agent Engineering speed-up in our AI software development case study.

Growth mindset in PM — useful as a culture, not a substitute for process

Magda and Peter Jaworowicz’s The Growth Mindset in Project Management applies Carol Dweck’s research to delivery teams. The distinction is familiar: fixed-mindset teams blame missed deadlines on impossible estimates, growth-mindset teams treat the miss as data and recalibrate.

Honest framing: there is not yet a peer-reviewed body of work that quantifies growth mindset’s ROI in software project metrics — cycle time, escaped defects, time-to-market. So treat it as a leading indicator of team adaptability, not a replacement for the rest of the discipline. Pair it with hard process: estimation reviews, retrospectives that change one specific practice each sprint, and a written team culture playbook.

What survived from February 2025 hype — and what didn’t

Mini case — how we restructured PM on a long-running EdTech build

On BrainCert, our long-running EdTech and virtual classroom platform, the situation in early 2025 was familiar to anyone running a mature product: estimates kept slipping a little, retros were polite, and stakeholders were tempted to throw an AI estimation tool at the problem.

We did the opposite. We rebuilt the estimation process around two rules. First, AI suggestions were allowed only as a bottom-up first draft for tasks with at least three historical analogues; everything else required a senior engineer to write the uncertainty bounds explicitly. Second, every retrospective had to ship one specific change before the next sprint — no “we’ll think about it”.

Within twelve weeks the variance between estimated and actual cycle time on planned work dropped meaningfully and the stakeholder conversations got easier — a pattern we have since reproduced across multiple clients. The general lesson: process discipline beats tooling spend almost every time.

A decision framework — pick what to adopt in five questions

1. Does it remove a real bottleneck? Audit where your PMs actually spend time. If status reports take 15% and risk reviews 5%, automating those is a clear win. If they spend 70% in stakeholder alignment, no AI tool fixes that.

2. Do you have the data the tool needs? Copilot for PM and AI estimators feed on historical metadata. New teams without that data should expect mediocre output.

3. What is the per-user, per-month all-in cost? Stack Microsoft 365 base + Copilot + meeting bot + Atlassian add-ons and the per-seat cost can pass $80/month before you blink.

4. Where does the conversation data go? If your industry is HIPAA, SOC 2, or EU privacy, validate every AI tool’s data residency and retention policy in writing before procurement.

5. Does it survive the “novel work” test? Run any AI estimation or summary against one current work item that is genuinely new for your team. If the answer is plausible-sounding nonsense, you have your verdict.

Five pitfalls we keep seeing in 2026 PM stacks

1. Treating AI productivity claims as data. Vendor case studies are marketing. Demand cycle-time, PR-review time, and escaped-defect data from comparable projects.

2. Buying SAFe ceremonies without buying SAFe discipline. Program Increments without honest dependency tracking are theatre. If you cannot point to one cross-team dependency that PI planning actually resolved, downsize the framework.

3. Stacking meeting bots without checking data residency. Audit every recording, transcript, and summary path against your compliance regime before you sign the renewal.

4. Letting AI estimates substitute for senior judgement. Anything truly new should be estimated by an engineer who has done something similar at least twice. AI is a starting point, not the answer.

5. Confusing certification with capability. A fresh PMI-ACP badge in 2027 will not, on its own, tell you whether the holder can run a working retrospective. Ask for the artefact, not the certificate.

KPIs to track on your PM and delivery process

Quality KPIs. Escaped-defect rate (target <3% of shipped tickets), PR review cycle time (target <48 hours median), and design-validation completion rate before build kickoff (target 100% for major epics).

Business KPIs. Estimated-vs-actual cycle time variance (target within ±15%), feature-to-revenue lag, and stakeholder satisfaction score per quarter.

Reliability KPIs. Sprint commit completion rate (target 80–90% — higher signals padding, lower signals chaos), team turnover (<15%/year for technical roles), and number of retrospective actions actually shipped per sprint (target ≥1).

When NOT to add another PM tool to your stack

If your team is under 12 people, your sprint commit rate is healthy, and your retros change behaviour, you do not need Copilot for PM, an integrated Atlassian stack, or a meeting-bot platform. The marginal complexity costs more than the time saved.

If you are pre-product-market fit, the highest-leverage PM investment is talking to users and shipping fast iterations. Heavy ceremony and AI tooling can wait until you have signal worth managing.

How to procure a partner with mature PM — the questions that work

Most procurement processes for software development partners over-index on price and capability slides and under-index on PM maturity. The result is the predictable failure mode: a partner that can write the code but cannot ship the project. Five practical questions cut through the noise.

Q1. “Show me a recent retrospective and the change it produced.” Mature teams can answer in two minutes. Immature teams pivot to credentials.

Q2. “What was your last estimate that missed by more than 30%, and what did the team learn?” The honest answer reveals more than any case study. Anyone who claims they have not missed an estimate by 30% in the last year is either too small to scale or lying.

Q3. “Walk me through how you instrument cycle time and PR review time.” If they cannot describe the dashboards or the source data, they are not measuring the work. Measurement absence is the strongest predictor of slipping deadlines.

Q4. “How do you decide when AI augmentation actually saves time?” The right answer mentions specific work types, not a percentage uplift. “30–40% faster on bottom-up planning for tasks with at least three historical analogues” is a real answer; “our team is 40% more productive” is marketing.

Q5. “If we have a HIPAA / SOC 2 / EU-residency requirement, walk me through your data flow for AI tools.” Concrete answers cite providers, BAAs, retention windows, and where audio actually goes. Vague answers are a fail.

Vendor red flags in the AI-PM era

Red flag 1 — “AI handles our PM.” Translation: the senior PM seat is unfilled and a tool is supposed to compensate. AI augments PMs; it does not replace them at any current state of the art.

Red flag 2 — status reports without raw data. Pretty PowerPoint summaries with no underlying ticket flow, code commit log, or QA cycle metrics tell you the partner is curating, not operating.

Red flag 3 — “everything is on track” for three sprints in a row. Real projects slip in small ways every sprint. A status report that never raises a risk is a status report that is not being read — or worse, not being written.

Red flag 4 — certifications instead of artefacts. A wall of PMP, PMI-ACP, CSM badges with no live retrospective example, no estimation worked-example, and no cycle-time data is decorative, not operational.

Red flag 5 — vague AI security stories. If they cannot tell you which AI providers their meeting bot routes audio through, where transcripts are stored, and for how long, do not turn it on for any conversation that touches PII, PHI, or commercially sensitive data.

FAQ

What to read next

Ready to sharpen your PM and delivery setup for 2026?

February 2025 made the picture clearer, not simpler. AI is real but slippery, scaled Agile is institutional but inconsistent, integrated tooling works at scale, the PMI just absorbed the Agile Alliance, and meeting bots are powerful but not free. The teams that win in 2026 are the ones with a written process, honest delivery data, AI used surgically rather than universally, and a partner whose PM maturity you can inspect on paper.

If you want a second opinion on where your delivery process is leaking time or money, that is exactly the conversation we have on a 30-minute scoping call. We bring our last six retros, our Agent Engineering benchmarks, and a written run-book proposal — you walk away with a prioritised list whether you hire us or not.