It\u2019s the day before the quarterly board pack is due. If you\u2019ve ever worked in a project portfolio office you know how that day goes.<\/p>\n
The numbers exist somewhere; the problem is getting them out, checking they\u2019re still valid and turning them into something the board will read.<\/p>\n
Most PMOs don\u2019t lack data, they\u2019re chronically short of the hours it takes to turn that data into something useful. The fix everyone looks for is another analyst and the answer is almost always the same: no headcount.<\/p>\n
We\u2019ve spent more than 20 years building project portfolio software for organisations. In that time this problem is the one we encounter more than any other. This post looks at where the PMO\u2019s time actually goes, what changes when an AI analyst takes on the collating and assembling work and the questions we think organisations should ask before letting AI anywhere near their portfolio data. This includes questions to ask us.<\/p>\n
Ask most PMO analysts about their job description and they\u2019ll talk about spotting the project that\u2019s off track, the next quarter\u2019s resourcing clash and the risks no-one has escalated yet.<\/p>\n
If you ask them what they actually spend their time doing, more often than not you\u2019ll get a very different answer. Instead of doing the strategic work they are employed to do they spend their time:<\/p>\n
Every reporting cycle someone manually pulls figures out of the system and rebuilds the same documents: the board pack, the portfolio summary, the status updates and the risk report. Even though the structure rarely changes from month to month the effort required to create it doesn\u2019t shrink.<\/p>\n
“Which projects are forecasting overspend?” sounds like a quick question. Answering it properly means filtering, cross-checking and confirming the figures are up to date. A handful of questions like that can take most of the day.<\/p>\n
Status updates go out of date. Risks never get logged. Benefits lose their link to the projects meant to deliver them. Data quality usually erodes quietly and isn\u2019t discovered until it\u2019s needed. This adds additional work to an already stretched team as they have to get the data into a state that\u2019s usable.<\/p>\n
None of the work mentioned above needs human involvement and this is where the opportunity to gain some of that badly needed time comes from. It\u2019s not replacing the analyst; it\u2019s giving them the time back to do the job they\u2019re employed to do.<\/p>\n
This is the gap we built Psonja to fill. She’s an AI analyst that works inside Psoda, she drafts reports from live portfolio data, answers questions in plain language and checks data quality.<\/p>\n
Having Psonja in the PMO changes things for everyone.<\/p>\n
Reporting stops being an all-consuming job. Board packs, portfolio summaries and status updates start as a draft built from the current state of the portfolio, so your team’s time goes into reviewing and refining instead of building it. Reporting season still arrives every year. It just doesn’t swallow the whole team for days at a time.<\/p>\n
The question “which projects are forecasting overspend?” becomes something you ask, not something you build. Psonja works directly with Psoda’s data. She can look up project information, move through the portfolio hierarchy and filter by type, keyword or attribute, so answers come from the live system rather than last month’s export. Follow-ups like “where are we overcommitted on resources next quarter?” or “what’s changed in the risk profile since the last board meeting?” take seconds rather than an afternoon.<\/p>\n
There’s also a library of prompts written and tested by PMO practitioners, so nobody needs to learn prompt engineering to get started.<\/p>\n
There are fewer last-minute requests for figures and fewer surprises. When data quality checks highlight a stale update or a missing risk, the project manager hears about it while it’s still a quick fix, not when it’s a finding in an audit report.<\/p>\n
Questions get answered faster. When a board member asks something that the pack didn’t cover, the PMO can find out properly and quickly instead of promising to come back next month.<\/p>\n
Data that stays in better shape between reporting cycles. Stale status updates, missing risks, outdated registers and orphaned benefits get flagged as they happen, rather than discovered all at once.<\/p>\n
In our experience, AI in the PMO rarely fails because the technology falls short. It stalls in the security review. Portfolio data is commercially sensitive, often politically sensitive and in the public sector it’s subject to frameworks that rightly ask hard questions. To address these concerns we made a few design decisions that we think matter more than any feature.<\/p>\n
You choose the AI.<\/strong> Psonja doesn’t come bundled with a model. You connect Psoda to Claude, Gemini, ChatGPT or Azure OpenAI using your organisation’s own API key and you choose the model. If your organisation has already approved Microsoft, Psonja can run on Azure OpenAI inside your own tenant. Your AI strategy stays yours and you can change model whenever you like.<\/p>\n Data goes only where you’ve already agreed.<\/strong> Each query\u2019s data passes to your chosen provider transiently under your existing agreement with them. It never goes through Psoda’s accounts. Your Psoda data stays in Psoda-owned infrastructure in New Zealand.<\/p>\n People see only what they’re already allowed to see.<\/strong> Psonja respects Psoda’s object-level permissions. If someone can’t see a project in Psoda, they can’t ask Psonja about it either. Administrators also decide who can use Psonja at all.<\/p>\n We tell you it’s AI.<\/strong> That sounds obvious, but it shapes everything. Psonja is AI-generated, like her work. When she introduces herself in our videos, the face and the voice are AI-generated too and she says so in the first few seconds. If an AI tool is going to work with your governance data, you should never have to guess whether you’re dealing with one.<\/p>\n Any honest description of an AI tool should cover its limits, so here are Psonja’s.<\/p>\n She won’t replace your judgement.<\/strong> Psonja drafts and the people decide. A board pack still needs someone who understands the politics, the context and what the board actually needs to hear.<\/p>\n She can’t fix data that isn’t there.<\/strong> If a project hasn’t been updated in two months, Psonja can tell you that, which is useful. She can’t tell you what’s really happening on that project. The data quality checks exist precisely because AI answers are only as good as the data underneath them.<\/p>\n She isn’t a general-purpose assistant.<\/strong> Psonja knows your portfolio, not your inbox or your file shares. That’s deliberate and it’s also why she works alongside tools like Copilot rather than replacing them (more on that below).<\/p>\n Whether you’re looking at Psonja or anything else, these are the questions we’d want answered before AI touches portfolio data. We’ve included our own answers so you can hold us to them.<\/p>\n <\/p>\nWhat it won’t do<\/strong><\/span><\/h2>\n
Questions worth asking any AI tool for your PMO<\/strong><\/span><\/h2>\n
\n\n
\n Question to ask<\/strong><\/td>\n Why it matters<\/strong><\/td>\n Psonja’s answer<\/strong><\/td>\n<\/tr>\n \n Whose AI model does it use, and can we choose?<\/td>\n A bundled model means accepting the vendor’s provider, terms and roadmap.<\/td>\n Yours: Claude, Gemini, ChatGPT or Azure OpenAI, on your own API key.<\/td>\n<\/tr>\n \n Where does our data go for each query?<\/td>\n “The cloud” isn’t an answer your security team can sign off.<\/td>\n Only to your chosen provider, transiently, per query, never through Psoda’s accounts.<\/td>\n<\/tr>\n \n Is our data used to train models?<\/td>\n Training on your data can expose it far beyond the original query.<\/td>\n Governed by your agreement with your provider. Standard enterprise API terms exclude training on your data.<\/td>\n<\/tr>\n \n Does it respect existing permissions?<\/td>\n An AI that can see everything turns every user into an administrator.<\/td>\n Yes. It uses Psoda’s object-level permissions, so users only get answers about what they can already see.<\/td>\n<\/tr>\n \n Can we control who uses it, and switch it off?<\/td>\n You need to roll out gradually and stop quickly if needed.<\/td>\n Administrators decide who has access, and it can be turned off per user at any time.<\/td>\n<\/tr>\n \n Does it work from live data or an export?<\/td>\n Answers from a stale export are confidently out of date.<\/td>\n Live Psoda data.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n