Hidden Costs and Lessons in AI Deployments

Key Takeaways:

• Bringing Salesforce’s Agentic Advisor into an existing CRM environment takes 1 to 2 years and carries a one-time implementation fee that a Salesforce reseller places in the $150,000 to $350,000 range, before license or AI usage pricing is added, according to RIABiz.
• SAP trimmed its 2026 operating profit outlook in July 2026 as AI-focused data acquisitions weighed on earnings, per Reuters.
• 55% of 400 organizations surveyed by Omdia spend 10 to 50 hours per month resolving inconsistent KPI or metric definitions, and 17% reported significant data-driven incidents, according to TechTarget’s report on that survey.
• Only 37% of leaders can attribute any positive EBIT impact to AI, and just 6% qualify as AI high performers, per McKinsey data cited by TechTarget.
• MIT’s NANDA research found purchased AI solutions succeed roughly 67% of the time versus internal builds succeeding about one-third as often.

In June 2026, a Salesforce reseller put a number on the thing every enterprise AI budget gets wrong. Franklin Tsung, chief growth advisor at the Salesforce reseller AppCrown, told RIABiz that bringing Agentic Advisor into an existing CRM environment takes 1 to 2 years and that the implementation fee alone, before license or AI usage pricing, lands in the $150,000 to $350,000 range. That is a reseller describing the product he sells, according to RIABiz. The license is the visible cost. The implementation, the data cleanup, and the ongoing human review are where the money actually goes.

This analysis covers what Salesforce and SAP deployments reveal about those hidden costs, what independent data says about the ROI gap, and how to budget for the line items vendors leave off the proposal.

Salesforce Deployment Costs: The Fees That Arrive After the License

Salesforce’s AI push has been expensive on both sides of the transaction. The company closed a $3.6-billion acquisition of Fin, an autonomous AI customer service platform, on June 15, 2026, and launched Agentic Advisor for financial advisors inside its Agentforce for Financial Services suite. Marc Benioff framed the Fin deal around “time to value,” a phrase that reads differently once you see the implementation numbers.

Tsung describes a cost structure with three layers. The first is the one-time implementation fee, which he places in the $150,000 to $350,000 range. The second is the Salesforce license. The third is AI usage-based pricing, which Tsung says “could hit millions if a firm is entirely relying on 100% Salesforce AI to run ALL workflows.” He also notes that the implementation process itself takes 1 to 2 years, which means the organization carries overhead and change-management costs for two budget cycles before the system is fully live.

The integration gap compounds the cost. Tsung says the agentic offerings “do not include turnkey infrastructure integrations with critical systems that support the wealth management industry” and describes the product as an “island” that needs “bridges and tunnels to bring data into Salesforce.” Every bridge and tunnel is custom integration work, and custom integration work is billed by the hour. Salesforce’s own press materials claim the platform “bypasses heavy, expensive customization cycles,” but that claim comes from the vendor, not from an independent deployment. The reseller who sells and integrates the product is the one describing the customization work as necessary.

Salesforce’s stock performance during this period shows the market was pricing in doubt about the AI transition. Shares lost 41% of their value year-to-date as of late June 2026, trading at $153, down 58% from a late-2024 peak, per RIABiz. That is context for why the company is pushing aggressive AI monetization: the revenue has to justify the valuation reset.

SAP Integration: Where the Modernization Bill Comes Due

SAP’s AI narrative runs through the same integration problem, and the company’s own financial disclosures show the cost. SAP trimmed its 2026 operating profit outlook in July 2026 as recent AI-focused data acquisitions weighed on earnings, according to Reuters. The acquisitions behind that trim include Reltio in March, then Dremio and Prior Labs in May, with the Dremio deal expected to close in Q3 2026 pending regulatory approval.

The logic for buying three data companies is straightforward, and SAP’s global head of AI go-to-market, Varun Thamba, states it directly: “70-80% of your data is SAP. Bring that 20%-ish over and then you can build something.” The cost of that strategy is that SAP is now integrating three acquisitions while asking customers to consolidate onto its platform. Thamba’s own description of the customer conversation has shifted from “how cool is your AI story” to “what value is your AI story bringing in 2026,” a change he attributes to rising token costs and the practical limits of agents in real processes.

SAP’s change-management costs are the underappreciated line. Thamba describes the problem of getting users to “bookmark so many URLs” and be trained on yet another AI interface, calling it a “big problem.” His answer is to embed AI where people already work, including making Joule available as a mobile app and inside Microsoft Teams, with the argument that “training is not required” and “change management is extremely minimized.” That is a vendor claim about a product direction, and it exists precisely because the alternative, training an entire workforce on a new AI front end, is expensive enough to threaten adoption.

Thamba also makes an argument against the multi-vendor approach many enterprises started with, comparing a fragmented model strategy to “having multiple credit cards” where “your loyalty is split.” The point is that consolidating on one platform reduces integration and contract costs. The trade-off is concentration risk: consolidating on SAP means the integration costs are SAP’s to set.

The Hidden Data Tax Nobody Puts on the Invoice

The largest hidden cost in enterprise AI is not a vendor fee at all. It is the labor spent making unreliable data usable enough for the model, and it never appears on the AI budget because it is distributed across teams that do not label it as AI work.

The numbers behind that hidden tax are specific. An Omdia survey of 400 organizations, reported by TechTarget in October 2026, found that 55% reported spending 10 to 50 hours per month resolving inconsistent KPI or metric definitions. Another 56% reported issues from AI systems operating on incomplete, inaccurate, or poorly governed data. Of those, 17% reported “significant” incidents that directly harmed business outcomes or financial performance, and 39% faced “moderate” disruptions requiring remediation.

Those hours are real money. An analyst who spends two days reconciling the definition of “revenue” across three systems records that time against a project, not against AI. An operations team double-checking model output charges it to operations. The rework from an AI decision made on bad data gets absorbed as cost of doing business. Each cost lands with the team that fixes it, and none of it rolls up into the total cost of the AI initiative.

The reason this matters more for Salesforce and SAP deployments than for a standalone chatbot is that both platforms sit on top of core systems of record. When an agent reads from CRM and writes to ERP, the quality of the identifiers and definitions across those systems determines whether the agent works. TechTarget’s analysis of the enterprise execution gap lists “unify primary data and context layer” as a core CIO requirement, noting that autonomous execution breaks down when core systems use conflicting names, IDs, and definitions. That reconciliation layer is a capital project, and enterprises routinely approve the far larger operating cost of paying employees to compensate for its absence instead.

The ROI Gap: Adoption Rises, EBIT Impact Stalls

The cost side is only half the story. The return side has barely moved despite a year of better models and broader deployment. McKinsey’s survey of 1,719 respondents across 97 countries found 44% of organizations scaling AI across the enterprise, up from 38% a year earlier, and 80% of employees saying AI improves their individual productivity. Yet only 37% of leaders can attribute any positive EBIT impact to AI, a figure that has not moved, and just 6% qualify as AI high performers, meaning AI accounts for at least 5% of their EBIT. That number is flat versus 2025, according to TechTarget’s September 2026 analysis of the McKinsey data.

The earlier MIT NANDA research, The GenAI Divide: State of AI in Business 2025, established the baseline. Based on 150 interviews with leaders, a survey of 350 employees, and analysis of 300 public AI deployments, it found that about 5% of AI pilot programs achieved rapid revenue acceleration while the vast majority stalled with little measurable P&L impact, as reported by Fortune. The report’s most actionable finding for buyers: purchasing AI tools from specialized vendors and building partnerships succeeded about 67% of the time, while internal builds succeeded only one-third as often. That is a direct argument against the instinct to build custom AI infrastructure in-house.

The pattern connecting both companies is that the bottleneck moved from answering to operating. A better model improves reasoning. It does not reconcile disconnected systems of record, resolve missing customer data, handle API failures, enforce permissions, or safely recover when a workflow drops halfway through. Those are the costs that decide whether a Salesforce or SAP deployment returns anything, and they are the costs least visible at approval time.

Cost Comparison: What the Line Items Actually Look Like

The table below separates the costs that appear on a vendor proposal from the ones that arrive later. Figures are drawn from the sources cited above; where a vendor does not publish pricing, the row says so rather than guessing.

The pattern in the table is that the vendor-controlled line items are bounded and quotable, while the organization-controlled line items are open-ended. A reseller can quote you a six-figure implementation range. Nobody can quote you the 10 to 50 hours per month your data team will spend reconciling definitions, because that cost depends on data quality the vendor never sees. For a deeper framework on the full cost stack, see our analysis of AI ROI measurement, which breaks down the human-in-the-loop and governance lines that dominate real budgets.

Lessons and a Budgeting Framework

The 2024-to-2026 record points to a few concrete practices that separate deployments that survive their second budget cycle from those that do not.

Budget the data remediation line explicitly. Add a “data remediation” line to every AI business case covering the time required to make the data foundation trustworthy for that use case. If it takes 40 hours of correction work to launch, that is an AI cost. The failure to name this category is the single most common reason AI business cases understate costs and overstate returns.

Buy rather than build, unless you have a structural reason not to. MIT’s finding that purchased solutions succeed about 67% of the time versus internal builds succeeding one-third as often is the strongest available evidence on this decision. The structural reasons to build, such as data residency, auditability, or IP control, are narrow. For most mid-market teams, a hybrid approach, a bought platform plus a custom retrieval layer on your own data, captures most of the value at a fraction of the cost, as we covered in our build-vs-buy cost comparison.

Measure process outcomes, not AI activity. Active users, prompt volume, and query counts do not connect to the balance sheet. The metrics that do are days sales outstanding, order-to-cash cycle time, cost per transaction, exception rate, and throughput per employee. If a metric cannot connect to revenue, cost, cash, risk, or capacity, it is not enterprise ROI.

Plan for a two-year runway, not a two-quarter one. Salesforce’s own reseller describes a 1 to 2 year implementation timeline, and the hidden data costs accumulate across the full lifecycle: sourcing and cleaning data before the project, building bespoke pipelines during development, and human review plus rework after deployment. The most expensive phase is the one after go-live, and it is the one most often left out of the plan.

Treat governance as a runtime cost, not a one-time review. Agents keep learning, adjusting, and drifting, and governance that runs as periodic review misses behavioral change between reviews. The organizations that capture value are the ones that redesign workflows around AI rather than bolting it onto existing processes. Nearly three-quarters of AI high performers redesign their business workflows around AI, compared with about one-quarter of everyone else, per the McKinsey data. That redesign work is a cost, and it is the cost that most reliably produces a return.

The through-line across Salesforce, SAP, and the broader enterprise data is that AI budgets fail on the same line items every time: integration, data remediation, change management, and ongoing human review. None of those are model costs. All of them are organizational costs, and all of them are visible before you sign if you know to look for them.

Related Reading

More in-depth coverage from this blog on closely related topics:

• AI Cybersecurity Threat Detection
• AI Agents for Business Automation

Sources and References

Sources cited while researching and writing this article:

• Salesforce tries to catch up on AI — and halt precipitous stock slide — with launch of 'Agentic Advisor' for RIAs, but experts are raising a litany of concerns from costs, to benefits and the hassle of implentation | RIABiz
• The hidden data costs missing from AI budgets
• Enterprise AI got better, but the ROI didn’t
• MIT report: 95% of generative AI pilots at companies are failing