EXECUTIVE BRIEFING
PSG EXECUTIVE BRIEFING · NO. 03
What AI actually costs.
The budget conversation nobody publishes: realistic price ranges for the four most common AI investments, the hidden costs that sink them, and the discipline that separates an investment from a write-off.
PLATINUM STRATEGY GROUP · 2026
WWW.PSG-INC.COM
Four common investments, priced like an operator.
Ranges below are for mid-market companies (roughly $10M–$500M revenue), all-in for the first year — software, implementation, and the internal time everyone forgets to count. Your position in the range depends almost entirely on one variable: the state of your data.
Copilot rollout
Commercial AI assistants across a department, with training and usage rules
$30K–$150K
Seat count and training depth. Licenses are the small part; adoption work is what makes seats produce anything.
Agent workflow
One automated process — invoice matching, report assembly, intake triage
$50K–$250K
Number of systems touched and exception complexity. A workflow with clean inputs costs a third of one that needs data repair first.
Private / offline LLM
Open-weight model behind your firewall with retrieval over your documents
$40K–$300K
Pilot on one workstation sits at the bottom; production with SSO, audit logs, and mid-class models at the top. Hardware is rarely the driver.
ML model in production
Forecasting or anomaly detection, backtested and operating on a cadence
$60K–$350K
Data assembly is half the project. Two clean years of history keeps you low; fragmented systems and moving definitions push you high.
The rule of thirds. In a well-run AI project, roughly a third of spend goes to the technology, a third to data readiness, and a third to adoption — training, process change, and the first ninety days of operation. Any proposal that prices only the first third is quoting you a down payment.
Six costs that never appear in the proposal.
01Data preparation
The largest and most consistently underquoted line. Deduplication, definition alignment, and access plumbing routinely double a quoted implementation. Ask every vendor: “what do you assume about our data, and what happens to price when that assumption fails?”
02Internal time
Your controller reviewing outputs, your ops lead in design sessions, your IT team on access and security. Plan for 0.5–1.5 internal FTEs during build. Projects “fully handled by the vendor” produce tools nobody trusts.
03Change management
Training, revised procedures, and the productivity dip while habits change. Typically 15–25% of project cost when done deliberately — and the whole project when skipped, because unused software is a 100% loss.
04Run cost drift
Per-token and per-seat pricing scales with success: the more your team uses it, the bigger the bill. Model your month-12 usage, not your month-1 pilot. Successful copilot rollouts commonly triple their run cost in year one.
05Maintenance & drift
Models degrade as the business changes; retrieval indexes go stale; prompts need tuning after every process change. Budget 15–20% of build cost annually. A system with no maintenance owner is a depreciating asset on day one.
06Exit costs
What does it cost to leave? Proprietary formats, non-portable configurations, and per-record export fees are all negotiable — before signature. The cheapest vendor with the most lock-in is rarely the cheapest vendor.
The comparison that matters. Weigh total year-one cost against the fully-loaded cost of the manual work it replaces — hours × loaded rate × error cost. When that math clears 2:1 in year one, fund it. When it needs a three-year horizon and heroic adoption assumptions to break even, it’s a science project wearing an ROI slide.
How disciplined buyers structure the spend.
Stage the commitment. Fund a bounded diagnostic or pilot first (10–20% of the full build), with defined success criteria and a decision date. Full funding follows evidence, not a demo. Any vendor who resists staging is telling you something.
Price the baseline before the tool. Measure the manual process now — hours, cycle time, error rate — so year-one value is measured, not asserted. This one week of work is what makes every later ROI conversation honest.
Budget all three thirds. Technology, data readiness, adoption — in writing, with owners. If the budget only covers software, the other two thirds will be paid anyway, unplanned, at premium rates, mid-project.
Set kill criteria up front. The cheapest AI project is the bad one you stop at week six. Agree in advance what evidence ends the project — and treat a clean early stop as a win, because it is one.
WHAT A REALISTIC FIRST-YEAR PORTFOLIO LOOKS LIKE
~$50K
Cautious: copilots in one department plus one scoped agent pilot. Proves value without betting the budget.
~$150K
Committed: one production agent workflow plus an ML forecast or anomaly model, measured against baseline.
~$350K+
Transformational: private LLM, two production workflows, and a reporting layer — sequenced over four quarters, not bought at once.
Want a real number for your situation?
A 30-minute diagnostic consultation — your use case, your data reality, and a costed range with the hidden thirds included, within 48 hours.