Get tomorrow's brief in your inbox
Today: Datasette's latest LLM plugin lets technical teams connect database agents to locally hosted models. Workday and Zuora are connecting usage-based billing with financial reporting, while Nvidia's leadership upheaval makes supplier continuity a practical concern for businesses buying AI computing.
Your AI workflow needs an exit plan, not just a subscription.
TechCrunch reports that privacy-focused assistant Ollie announced its shutdown and Character.ai filed for bankruptcy. It also cites S&P Global Market Intelligence's finding that about 42% of corporate AI initiatives are ultimately abandoned, a broader measure than startup failures. For a small business, the practical lesson is straightforward: a useful demo does not guarantee a dependable long-term service. Source
Before making an AI tool essential to daily operations, test whether you can leave it. Check whether you can export your documents, conversation history, workflow definitions, and other business records in usable formats. Keep your operating procedures outside the tool, identify who owns the account, and document a manual fallback for customer-facing work. These are recommended purchasing checks, not requirements reported by the article.
Business impact: Portability can reduce the disruption of a shutdown, product retirement, or forced migration. You do not need two paid subscriptions for every task, but you should know how to complete tomorrow's quotes, customer replies, or reports without your preferred assistant. Make that continuity check part of the buying decision alongside output quality and subscription cost.
datasette-llm 0.1a11 adds support for OpenAI-compatible endpoints, including models served through LM Studio, Ollama, or llama.cpp. This is useful for developers and technically capable analysts who want to use Datasette plugins with a model endpoint they control, rather than being tied to one hosted provider. SourceWhat to do: Have your developer test Datasette Agent against a small, non-sensitive database and a local model server. Ask representative business questions, then compare the answers with verified queries before expanding access.
Cost and impact: Confirm the chosen model's licensing and your hardware requirements before deployment. Local hosting changes the cost structure, but it does not establish that the workflow is cheaper, more accurate, or fully private.
datasette-vega 0.7 release updates compatibility with recent Datasette alphas and improves its operation with Datasette Lite; it renders Vega-Lite charts from tables or custom queries. For an existing Datasette user, this offers a way to visualize database results alongside an AI-assisted analysis workflow, although the charting plugin itself is not an AI feature. SourceWhat to do: Test it on one recurring report, such as sales by month or orders by category, and check the chart against the underlying table. The opportunity is easier inspection of results, not a substitute for validating the analysis.
What to do: If you are building an agent around sensitive business data, borrow the design principles rather than assuming this particular application is available to buy. Define access permissions, isolate analysis sessions, retain interaction traces, and maintain evaluation tests when changing infrastructure.
Business impact: Managed infrastructure can shift some operating work to the platform provider, letting your team concentrate on its application. It does not remove your responsibility to test outputs, control access, or verify that the deployment meets your data-protection obligations.
A practical pilot: Choose one bounded analysis task with a known answer and an accountable reviewer. Measure correctness, review effort, and operating cost before giving the agent broader data access. The clinical project supports an architecture discussion, not a claim that an agent can safely make unsupervised medical or business decisions.
Who should investigate: Businesses already using Workday that need usage-based or hybrid billing, including companies selling AI-enabled services. This is a finance-integration development, not evidence that every small business needs a new billing platform.
What to do: Request current pricing and an implementation estimate, then ask for a demonstration using your actual charging rules. Include usage measurement, billing corrections, and reconciliation in the test, rather than judging the system only on how quickly it produces an invoice.
Business impact: Connecting usage, charges, and revenue could reduce manual reconciliation. The buying decision should compare that potential saving with integration work and ongoing administration; introducing more flexible pricing also creates more things your finance team must verify.
Separately, Megaspeed told CNBC that it had opened an internal investigation into reporting about its business and supply chain, while saying those reports were unsupported by verifiable evidence. Allegations concerning chip diversion remain allegations, not established findings. Source
What to do: If you rent GPU capacity, ask your provider to confirm availability, applicable export restrictions, and its contingency arrangements. Do not change providers solely because of leadership headlines. The business issue is whether your contracted service remains dependable.
What to do: If a server refresh is already planned, request a comparison using your database, application, or agent-orchestration workload. For cloud deployments, confirm the relevant instance's availability and current price before making a migration plan.
Business impact: Better CPU efficiency could improve infrastructure economics, but measure total cost, including migration and administration. A new processor announcement alone is not a reason to replace working equipment.
Run a 30-minute AI continuity drill. Pick your most important AI-assisted task, export the supporting material, and ask a colleague to complete it without the usual tool. Record what blocks them and update your fallback procedure. This recommendation follows the product-closure risks described in TechCrunch's AI graveyard report.
Create a five-question database-agent test. Before trying Datasette with a local model, write five questions whose answers you already know, including one the database cannot answer. Score factual accuracy and whether the agent admits insufficient information; this gives you a practical acceptance test before involving business users. Related release
The useful thread is control: choosing model endpoints, isolating agent sessions, connecting usage to billing, and keeping workflows portable. This week, prioritize a small test with measurable results over a broad rollout, and verify availability, pricing, and contractual commitments before changing infrastructure or suppliers.